modeling and simulation of piezocomposites
TRANSCRIPT
Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232
www.elsevier.com/locate/cma
Modeling and simulation of piezocomposites
Jacob Fish *, Wen Chen
Department of Mechanical, Aerospace and Nuclear Engineering, Rensselaer Polytechnic Institute, 4042 Johnnson Engineering Center,
Troy, NY 12180-3590, USA
Received 12 May 2003; accepted 12 May 2003
Abstract
Solution procedures for large-scale transient analysis of piezocomposites are developed. Computational efficiency of
unconditionally stable and conditionally stable implicit partitioned schemes as well as explicit partitioned schemes is
investigated in the context of multilevel solution methods applied to separate field equations.
� 2003 Elsevier B.V. All rights reserved.
1. Introduction
Multiphysics modeling and simulation has recently gained significant attention in computational science
and engineering community. Example problems falling into this category are coupled mechanical, thermal,
electrical, diffusion-reaction, fluid-structure interaction problems. In multiphysics problems, typically awide range of interacting spatial and temporal scales are involved resulting in large coupled time-dependent
system of equations.
Generally, there are three approaches for simulation of coupled systems [1]: field elimination, monolithic
and partitioned schemes. Field elimination is restricted to linear (or linearized) problems that permit effi-
cient decoupling. The monolithic scheme is in general computationally challenging, economically subop-
timal and software-wise unmanageable [2]. On the other hand, the partitioned treatment, which is the focus
of the present paper, is computationally efficient, flexible and software reusable.
In partitioned schemes, computational efficiency is counteracted by the fact that satisfactory numericalstability is hard to achieve. In fact the practical feasibility of the partitioned approach hinges entirely on the
stability analysis [1]. Stability analysis by standard Fourier techniques using a scalar test equation is not
generally possible because modes of individual subsystems are not modes of the coupled problem. More-
over, partitioned procedures are in general less accurate than their underlying subsystem time integrators.
In principle, higher order accuracy can be recovered by iterating the solution between the fields at each time
step [3]. However, the computational cost of interfield iterations may overshadow the gains from reduced
time steping [1].
* Corresponding author. Tel.: +1-518-276-6191; fax: +1-518-276-4833.
E-mail address: [email protected] (J. Fish).
0045-7825/03/$ - see front matter � 2003 Elsevier B.V. All rights reserved.
doi:10.1016/S0045-7825(03)00343-8
3212 J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232
In this paper we focus on the problem of piezoelectricity in periodic heterogeneous media. We start byformulating the monolithic equations and study stability of implicit time integrators (the Newmark
method) applied to the monolithic equations. It is demonstrated that taking the integration parameters
which render the Newmark method unconditionally stable for conventional structural dynamics problems
do not guarantee unconditional stability of the monolithic system. This is due to the fact that in structural
dynamics problems, the global mass matrix is positive definite and the stiffness matrix is at least positive
semi-definite, whereas for piezocomposites the global stiffness matrix of the monolithic scheme is indefinite,
and thus conventional implicit integrators directly applied to the monolithic system may not be stable.
From the monolithic equations, we construct the implicit partitioned schemes based on the block Gauss–Seidel method. It is shown that the partitioned scheme constructed directly from the monolithic equations
does not guarantee unconditional stability. Unconditionally stable partitioned scheme can be obtained by
augmenting the original equations of motion and subsequently splitting the field equations along the lines
of block Gauss–Seidel method.
Computational efficiency of solving the problem of piezoelectricity is the main focus of the present
paper. While partitioned approach reduces the solution of coupled physics problem into a sequence of
single field problems, the computational cost could be still insurmountable due to the existence of multiple
spatial scales within each physical process. For this type of problems the direct method may not compu-tationally optimal, while various multigrid methods though very efficient for differential equations with
constant coefficients, exhibit poor rates of convergence when applied to differential equations with rapidly
oscillating coefficients [4,5]. The representative volume element (RVE)-based multilevel method developed
by the authors has been shown to possess high rates of convergence for single-field structural dynamics
problems in periodic heterogeneous media [6]. In the present paper, we extend the framework of the RVE-
based multilevel method to transient analysis of piezocomposites.
Convergence properties of the RVE-based multilevel method are studied in the context of uncondi-
tionally stable and conditionally stable implicit and explicit partitioned schemes. We show that the rate ofconvergence of the RVE-based multilevel method for conditionally stable implicit and explicit partitioned
schemes is higher than for unconditionally stable scheme. This is in part due to the fact that the eigenspace
of the stabilized matrix cannot be reproduced by a linear combination of the RVEs. For high frequency
applications, the conditionally stable implicit and explicit partitioned schemes are advantageous, since the
choice of time step is governed by accuracy considerations, whereas the unconditionally stable schemes
might be better suited for low frequency applications.
2. Finite element formulation of piezoelectricity
2.1. Governing equations of piezoelectricity
Let X denote the domain of interest, and C the boundary of X. Consider the governing equations of
linear piezoelectricity, which include:
ii(i) Mechanical equilibrium
rij;j þ fi ¼ q€uui; ð1Þi(ii) Maxwell�s equation of equilibrium (assumed to be quasi-static and decoupled from mechanical fields)
Di;i ¼ 0; ð2Þ(iii) Constitutive equations
rij ¼ CijklSkl � ekijEk; Di ¼ eiklSkl þ eikEk; ð3Þ
J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232 3213
(iv) Strain–displacement relation and the curl-free electric field
Sij ¼ 12ðui;j þ uj;iÞ; Ei ¼ �/;i; ð4Þ
where r, S, E and D are the stress, strain tensors, electric field and electric displacement vectors, respec-tively; C , e, and e are the elastic tensor at constant electric field, dielectric tensor at constant strain field and
the piezoelectric tensor, respectively; u and / denote the displacement vector and electric potential (volt-
age), respectively; q is the mass density and f the body force. The superposed dot and comma denote
temporal and spatial differentiation, respectively. Since the piezoelectric tensor e is a third-rank tensor,
piezoelectric coupling exists in intrinsically anisotropic materials only [7].
To complete the description of the problem, the boundary conditions are prescribed as
ui ¼ �uui on Cu and / ¼ �// on C/; ð5Þ
rijnj ¼ T i on Cr and Dini ¼ �qq on Cq; ð6Þ
Cu [ Cr ¼ C; Cu \ Cr ¼ 0; C/ [ Cq ¼ C; C/ \ Cq ¼ 0; ð7Þwhere �uu and T are the prescribed mechanical displacement and surface traction, respectively; �// and �qq are
the prescribed electric potential and surface charge, respectively; n denotes the outward unit normal vector.
2.2. The weak formulation and semi-discrete equations of motion
From the governing equations (1)–(4) and boundary conditions (5)–(7), we establish the weak formu-
lation of the problem:
For t 2 ð0; T �, find uiðx; tÞ 2 V �uum and /ðxÞ 2 V �//
e , such thatZX½q€uui � ðrij;j þ fiÞ�wi dX ¼ 0 8wi 2 V 0
m ; ð8Þ
ZXDi;iWdX ¼ 0 8W 2 V 0
e ; ð9Þ
where
V �uum ¼ wðxÞ 2 H 1ðXÞ;w
n¼ �uu on Cu
o; ð10Þ
V�//
e ¼ /ðxÞ 2 H 1ðXÞ;/n
¼ �// on C/
o; ð11Þ
V 0m ¼ V �uu
m
����uu¼0
; V 0e ¼ V
�//e
����//¼0
; ð12Þ
H 1ðXÞ ¼ w�
¼ wðxÞ; x 2 Xjw;rwi 2 L2ðXÞ�
ð13Þ
with L2ðXÞ denoting the set of square-integrable functions over X.Integrating Eqs. (8) and (9) by parts and substituting the constitutive relations (3), the kinematics re-
lation (4) and the natural boundary conditions (6) into the resulting equations, and exploiting the symmetry
of the elastic tensor, we haveZXqwi€uui dXþ
ZXSijðwÞ CijklSklðuÞ
�þ ekij/;k
�dX ¼
ZXfiwi dXþ
ZCr
T iwi ds; ð14Þ
3214 J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232
ZXW;i eiklSklðuÞ
�� eik/;k
�dX ¼
ZCq
�qqWds; ð15Þ
where SðwÞ ¼ rsw is the symmetric gradient of w.Subsequently the matrix notation is adopted. We employ finite element discretization with element shape
functions denoted as Nu and N/
u ¼ NuUe; / ¼ N/U
e; ð16Þwhere Ue and Ue are element nodal values. Differentiating Eq. (16) yields the element strain and electricfield vectors
S ¼ BuUe; E ¼ �r/ ¼ �B/U
e: ð17ÞInserting Eqs. (16) and (17) into (14) and (15) we obtain the semi-discrete equations of motion of piezo-
electricity
Muu€UU þ KuuU þ Ku/U ¼ F; K/uU � K//U ¼ Q; ð18Þ
where the assembly operators are defined in a usual manner:
Muu ¼Ne
e¼1
ZXe
qNTuNu dXe consistent mass matrix;
Kuu ¼Ne
e¼1
ZXe
BTuCBu dXe mechanical stiffness matrix;
Ku/ ¼Ne
e¼1
ZXe
BTu e
TB/ dXe; K/u ¼Ne
e¼1
ZXe
BT/eBu dXe piezoelectric stiffness matrices;
K// ¼Ne
e¼1
ZXe
BT/eB/ dXe dielectric stiffness matrix;
F ¼Ne
e¼1
ZXe
NTu f dXe þ
Sre
e¼1
ZCer
NTuT ds; Q ¼
Sqe
e¼1
ZCeq
NT/�qqds forcing vectors;
with Ne and Sre , S
qe denoting the total number of elements and the number of elements with prescribed
boundary conditions, respectively.
3. Field elimination and monolithic solution schemes
3.1. Static condensation
Suitably grounding the structure by specifying one or more nodal values of the potential renders K//
non-singular. From the second equation in (18), we have
U ¼ K�1// KT
u/U�
�Q�: ð19Þ
Substituting the above equation into the first equation in (18) yields
Muu€UU þ eKKU ¼ eFF ; ð20Þ
J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232 3215
whereeKK ¼ Kuu þ Ku/K�1//K
Tu/;
eFF ¼ F þ Ku/K�1//Q: ð21Þ
Eq. (20) is obtained by field elimination or static condensation of the electric field. It can be observed
that eKK and Muu are positive definite, and Eq. (20) can be easily solved by using either implicit or explicit
time integrator. This approach has been adopted in [8,9]. While relatively easy to implement for small
problems, static condensation suffers a number of drawbacks. First, the condensed stiffness matrix eKK is
denser than Kuu, and second, it destroys spectral properties of Kuu, which could be taken advantage of in
designing efficient multilevel iterative solvers.
3.2. Monolithic scheme
Exploiting K/u ¼ KTu/, we can rearrange Eq. (18) in the monolithic form:
M€dd þ Kd ¼ F; ð22Þwhere
M ¼ Muu 0
0 0
� ; K ¼ Kuu Ku/
KTu/ �K//
� ; d ¼ U
U
� ; F ¼ F
Q
� : ð23Þ
It can be seen that M is positive semi-definite and K indefinite. The monolithic scheme for piezoelec-
tricity has been adopted by a number of authors [10–12]. However, prior to applying time integration
scheme with proven stability behavior to the monolithic equation (22) it is necessary to examine its suit-ability for problems of piezoelectricity. We focus on Newmark integrator
dnþ1 ¼ dn þ Dtvn þ Dt2 12
�� b
�an þ banþ1
�; ð24Þ
vnþ1 ¼ vn þ Dt ð1½ � cÞan þ canþ1�; ð25Þwhere dn, vn and an are the displacement, velocity and acceleration vectors at time tn, respectively; Dt is thetime step size; b and c are parameters which determine the stability and accuracy of the Newmark method.
For structural dynamics the Newmark scheme is unconditionally stable for 2bP cP 1=2.The stability of the Newmark method can be investigated by using the energy method (cf. [13,14]).
Eqs. (22) (with F ¼ 0), (24) and (25) can be combined to form the following identity:
aTnþ1Banþ1 þ vTnþ1Kvnþ1 ¼ aTnBan þ vTnKvn � ð2c� 1Þ½an�TB½an�; ð26Þ
where
B ¼ M þ Dt2 b
� 12c�K ; ½an� ¼ anþ1 � an: ð27Þ
From the above two equations, it can be observed that if cP 1=2 and B is positive definite, we have
aTnþ1Banþ1 þ vTnþ1Kvnþ1 6 aTnBan þ vTnKvn ð28Þ
which implies
aTnBan þ vTnKvn 6 aT0Ba0 þ vT0Kv0; n ¼ 1; 2; 3; . . . ð29ÞTherefore, an and vn are bounded. If K is non-singular, from Eq. (22) (with F ¼ 0), we have kdnk6
kK�1kkMkkank, and then dn is bounded. For M and K to be positive definite and at least positive semi-
definite respectively, as is the case of elasto-dynamics, the matrix B is positive definite for 2bP c, which is
the well-known condition for the Newmark scheme to be unconditionally stable. In the case of piezo-
electricity M and K are positive semi-definite and indefinite, respectively, and thus the condition of 2bP c
3216 J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232
does not guarantee positive-definiteness of B. Nevertheless, in case of the average acceleration method (ortrapezoidal rule) with b ¼ 1=4 and c ¼ 1=2, the monolithic scheme applied to linear piezoelectricity is
unconditionally stable, in which case Eqs. (26) and (29) reduce to
aTnþ1Manþ1 þ vTnþ1Kvnþ1 ¼ aTnBan þ vTnKvn; ð30Þ
aTnMan þ vTnKvn ¼ aT0Ma0 þ vT0Kv0; n ¼ 1; 2; 3; . . . ð31Þ
4. Partitioned schemes
Applying the trapezoidal rule to the monolithic equation (22) yields the implicit unconditionally stable
integration scheme, as described in the previous section:
bKKdnþ1 ¼ bFF nþ1; vnþ1 ¼2
Dtðdnþ1 � dnÞ � vn; ð32Þ
where
bKK ¼ K þ 4
Dt2M ; bFF nþ1 ¼
4
Dt2M
�� K
dn þ
4
DtMvn þ Fn þ Fnþ1 ð33Þ
are the effective stiffness matrix and load vectors, respectively. First, we consider implicit partitioned
schemes directly constructed from the above monolithic scheme and show that the resulting partitionedscheme is not unconditionally stable. Unconditionally stable implicit partitioned scheme is then obtained
via augmentation. An explicit partitioned scheme is also constructed from the explicit integration of
mechanical equation.
4.1. Partitioning directly from the monolithic scheme
Partitioned schemes can be constructed directly from the monolithic scheme by splitting the effective
stiffness matrix. The first equation of the monolithic scheme (32) can be written as
bKK uu Ku/
KTu/ �K//
" #Unþ1
Unþ1
� ¼
bFF unþ1
Qnþ1
" #; ð34Þ
where
bKK uu ¼ Kuu þ4
Dt2Muu; bFF u
nþ1 ¼ Muu4
Dt2Un
�þ 4
Dt_UUn þ €UUn
þ Fnþ1: ð35Þ
We start by considering a partitioned scheme where the electric field is predicted first. Splitting the
coefficient matrix of (34) based on the block Gauss–Seidel method yields
bKK uu 0
KTu/ �K//
" #Unþ1
Unþ1
� ¼ � 0 Ku/
0 0
� Unþ1
Unþ1
� þ
bFF unþ1
Qnþ1
" #: ð36Þ
Utilizing the last-step-value predictor on the right hand side of (36), we havebKK uu 0
KTu/ �K//
" #Unþ1
Unþ1
� ¼ � 0 Ku/
0 0
� Un
Un
� þ
bFF unþ1
Qnþ1
" #: ð37Þ
J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232 3217
Expanding the above equation yields the electric field predicted partitioned scheme:
bKK uuUnþ1 ¼ bFF unþ1 � Ku/Un; K//Unþ1 ¼ �Qnþ1 þ KT
u/Unþ1: ð38Þ
Alternatively, we can predict the mechanical field. The monolithic equation (34) can be written as
�K// KTu/
Ku/bKK uu
" #Unþ1
Unþ1
� ¼ Qnþ1bFF u
nþ1
" #: ð39Þ
Splitting the coefficient matrix of the above equation based on the block Gauss–Seidel method yields
�K// 0
Ku/bKK uu
� Unþ1
Unþ1
� ¼ � 0 KT
u/
0 0
� Unþ1
Unþ1
� þ Qnþ1bFF u
nþ1
" #: ð40Þ
Utilizing the last-step-value predictor and expanding the resulting equation lead to the mechanical fieldpredicted partitioned scheme
K//Unþ1 ¼ �Qnþ1 þ KTu/Un; bKK uuUnþ1 ¼ bFF u
nþ1 � Ku/Unþ1: ð41Þ
Both partitioned schemes solve the two field equations independently at each time step.
4.1.1. Stability analysis of the implicit partitioned schemes
Consider the partitioned procedures (38) and (41) which predict the electric field and mechanical field,
respectively. Both partitioned procedures compute the velocity and acceleration fields based on the fol-
lowing equations:
_UUnþ1 ¼ _UUn þDt2
€UUn
hþM�1
uu ðFnþ1 � KuuUnþ1 � Ku/Unþ1Þi; ð42Þ
€UUnþ1 ¼ M�1uu ðFnþ1 � KuuUnþ1 � Ku/Unþ1Þ: ð43Þ
For the purpose of stability analysis, it suffices to consider the force-free case, i.e., Fnþ1 ¼ Qnþ1 ¼ 0.
Stability analysis is carried out using standard procedures [15,16]. Non-trivial solutions of the partitioned
procedures are sought in the form
Unþ1 ¼ kUn; _UUnþ1 ¼ k _UUn; €UUnþ1 ¼ k €UUn; Unþ1 ¼ kUn: ð44ÞStability requires jkj6 1. When the modulus is unity, the root must be a simple one [17]. Replacing k
by ð1þ zÞ=ð1� zÞ, stability condition reduces to
ReðzÞ6 0: ð45ÞFirst, we consider the mechanical field predicted partitioned procedure. Substituting Eq. (44) into (41)–
(43) yields
� 1þ z1� z
K// KTu/
Ku/ Kuu þ4z2
Dt2Muu
264375 Un
Un
� ¼ 0: ð46Þ
The characteristic polynomial associated with (46) is
z34
Dt2Muu
���� þ z24
Dt2Muu þ z Kuu
�� Ku/K
�1//K
Tu/
�þ Kuu þ Ku/K
�1//K
Tu/
���� ¼ 0; ð47Þ
3218 J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232
where j � j denotes the matrix determinant. The necessary condition for stability is that all coefficient ma-
trices of the polynomial in z should be positive definite (the first part of Routh–Hurwitz conditions [17,18]).
Since Muu and Kuu þ Ku/K�1//K
Tu/ are positive definite, but Kuu � Ku/K
�1//K
Tu/ is indefinite, the necessary
condition for stability cannot be satisfied.
Next, we investigate the electric field predicted partitioned scheme. Substituting Eq. (44) into (38), (42)
and (43) yields
K// �KTu/
ð1� z2ÞKu/ Kuu þ4z2
Dt2Muu
24 35 Un
Un
� ¼ 0: ð48Þ
The characteristic polynomial of (48) is
Muu
����� � Dt2
4Ku/K
�1//K
Tu/
z2 þ Dt2
4Kuu
�þ Ku/K
�1//K
Tu/
����� ¼ 0: ð49Þ
To investigate the stability properties of the above equation it is convenient to utilize the theorem of
Bellman [18], which states: If A, B and C are non-negative definite, and either A or C positive definite, then
z2A�� þ 2zB þ C
�� ¼ 0 ð50Þ
has no roots with positive real parts. The proof is given in [18].In view of the above theorem, it is obvious that the necessary and sufficient condition for stability of the
electric field predicted partitioned scheme is that the matrix Muu � Dt2
4Ku/K
�1//K
Tu/ is non-negative definite,
which can be achieved with sufficiently small Dt. Therefore, the scheme is conditionally stable.
We proceed to investigate the critical time step of the above scheme. Consider the eigenvalue problem
Muu
�� Dt2
4Ku/K
�1//K
Tu/
x ¼ cx: ð51Þ
The necessary and sufficient condition for stability becomes
cP 0: ð52Þ
From the generalized eigenvalue problem
Ku/K�1//K
Tu/
h� ðximÞ2Muu
ix ¼ 0; ð53Þ
we have
M�1uu Ku/K
�1//K
Tu/x ¼ ðximÞ2x ð54Þ
Premultiplying (51) by M�1uu yields
x� Dt2
4M�1
uu Ku/K�1//K
Tu/x ¼ cM�1
uu x: ð55Þ
Substituting (54) into the above equation yields
cM�1uu x ¼ 1
h� ðximDtÞ2=4
ix: ð56Þ
If c ¼ 0, we have from the above equation
Dt ¼ 2=xim: ð57Þ
J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232 3219
If c 6¼ 0, we have from (56)
M�1uu x ¼
1� ðximDtÞ2=4h i
cx: ð58Þ
The positive definiteness of M�1uu and Eq. (52) lead to
1� ðximDtÞ2=4 > 0: ð59Þ
From Eqs. (57) and (59), it follows that
Dt6 2=xim: ð60Þ
The above stability condition must be satisfied for each system mode. Consequently, the highest fre-
quency ximn is critical and must satisfy (60). Therefore, the stability condition becomes
Dt6Dtimcr ¼ 2=ximn : ð61Þ
4.1.2. Consistency and accuracy of the electric field predicted implicit scheme
By employing the equilibrium equation at tn to eliminate €UUn, we can express the electric field predicted
partitioned scheme (38), (42) and (43) in the following form
dnþ1 ¼ Bdn þ Pn; ð62Þ
where B ¼ B�11 B2,
B1 ¼
bKK uu 0 0
KTu/ 0 �K//
Dt2M�1
uu Kuu IDt2M�1
uu Ku/
266664377775; B2 ¼
4
Dt2Muu � Kuu
4
DtMuu �2Ku/
0 0 0
�Dt2M�1
uu Kuu I �Dt2M�1
uu Ku/
2666664
3777775;
dn ¼
Un
_UUn
Un
26643775; Pn ¼ B�1
1
Fn þ Fnþ1
Qnþ1
Dt2M�1
uu ðFn þ Fnþ1Þ
266664377775:
ð63Þ
Inserting the exact solution dðtnÞ into (62), we have
dðtnþ1Þ ¼ BdðtnÞ þ Pn þRðtnÞ; ð64Þ
where RðtnÞ ¼ ½RTu ðtnÞ _RRT
u ðtuÞ RT/ðtnÞ �
Tis the local truncation error. Inserting the following Taylor
expansions:
dðtnþ1Þ ¼ dðtnÞ þ Dt _ddðtnÞ þDt2
2€ddðtnÞ þ
Dt3
6
vdðtnÞ þOðDt4Þ; ð65Þ
Fnþ1 ¼ F þ Dt _FFn þDt2
2€FFn þ
Dt3
6
vFn þOðDt4Þ; ð66Þ
3220 J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232
Qnþ1 ¼ Qn þ Dt _QQn þDt2
2€QQn þ
Dt3
6
vQn þOðDt4Þ ð67Þ
into Eq. (64) and expanding the resulting equation, it can be shown that
RðtnÞ � OðDt3Þ and RðtnÞ ! 0 as Dt ! 0: ð68ÞTherefore, the electric field predicted partitioned scheme is consistent and second-order accurate.
4.2. Unconditionally stable implicit partitioned scheme
4.2.1. Stabilization by semi-algebraic augmentation
The stabilization is carried out by a semi-algebraic augmentation technique proposed by Farhat et al.
[15] for thermoelastic problems.
First, the structural equation in (18) is integrated using the trapezoidal rule:
_UUnþ1 ¼ _UUn þDt2ð €UUn þ €UUnþ1Þ ¼ _UUn þ
Dt2
€UUn
hþM�1
uu ðFnþ1 � KuuUnþ1 � Ku/Unþ1Þi; ð69Þ
Unþ1 ¼ Un þ Dt _UUn þDt2
4ð €UUn þ €UUnþ1Þ ¼ Un þ Dt _UUn
Dt2
4€UUn
hþM�1
uu ðFnþ1 � KuuUnþ1 � Ku/Unþ1Þi:
ð70ÞConsider the electric equation
KTu/Unþ1 � K//Unþ1 ¼ Qnþ1: ð71Þ
Substituting Eq. (70) into the above equation yields
bKK /uUnþ1 þ bKK //Unþ1 ¼ bFF/nþ1; ð72Þ
where
bKK /u ¼Dt2
4KT
u/M�1uu Kuu; bKK // ¼ K// þ Dt2
4KT
u/M�1uu Ku/;
bFF/nþ1 ¼
Dt2
4KT
u/M�1uu Fnþ1 þ KT
u/ Un
�þ Dt _UUn þ
Dt2
4€UUn
�Qnþ1:
ð73Þ
Rearranging (70) yields
bKK uuUnþ1 þ Ku/Unþ1 ¼ bFF unþ1; ð74Þ
where bKK uu and bFF unþ1 are given in (35). Eqs. (72) and (74) can be written as
bKK //bKK /u
Ku/bKK uu
" #Unþ1
Unþ1
� ¼
bFF/nþ1bFF unþ1
" #: ð75Þ
Splitting the coefficient matrix in (75) using the block Gauss–Seidel method and last-step-value pre-
dictor, we have
bKK // 0
Ku/bKK uu
" #Unþ1
Unþ1
� ¼ � 0 bKK /u
0 0
� Un
Un
� þ
bFF/nþ1bFF unþ1
" #: ð76Þ
J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232 3221
Expanding the above equation and using (69) yield the following partitioned algorithm:
1. Predict the mechanical field
Upnþ1 ¼ Un: ð77Þ
2. Solve for the electric field
bKK //Unþ1 ¼ bFF/nþ1 � bKK /uU
pnþ1: ð78Þ
3. Correct the mechanical field
bKK uuUnþ1 ¼ bFF unþ1 � Ku/Unþ1: ð79Þ
4. Compute velocity and acceleration fields
_UUnþ1 ¼ _UUn þDt2
€UUn
hþM�1
uu ðFnþ1 � KuuUnþ1 � Ku/Unþ1Þi; ð80Þ
€UUnþ1 ¼ M�1uu ðFnþ1 � KuuUnþ1 � Ku/Unþ1Þ: ð81Þ
4.2.2. Stability and accuracy of the stabilized implicit scheme
4.2.2.1. Stability. We consider the force-free case, i.e. Fnþ1 ¼ Qnþ1 ¼ 0. The stability of the proposed
partitioned scheme can be examined by seeking a non-trivial solution in the form of (44). Substituting (77)
into (78), and (44) into (78)–(81), we have after some algebraic manipulations
Kuu þ4z2
Dt2Muu Ku/
Dt2
2ð1� z2ÞVKuu � zð1� zÞK�1
//KTu/ zðzþ 1ÞI þ Dt2
2VK u/
264375 Un
Un
� ¼ 0; ð82Þ
where
V ¼ K�1//K
Tu/M
�1uu : ð83Þ
The characteristic polynomial associated with (82) is
det Muuz3�
þMuuz2 þDt2z4
Kuu
�þ Ku/K
�1//K
Tu/ I
�þ Dt2
2M�1
uu Kuu
þ Dt2
4Kuu
�þ Ku/K
�1//K
Tu/
��¼ 0:
ð84ÞRecall that Muu, Kuu and K// are positive definite; Ku/K
�1//K
Tu/ is positive definite provided that Ku/ has
full column rank and positive semi-definite if Ku/ is column rank deficient. In any case, all coefficient
matrices of the characteristic polynomial (84) are positive definite. Therefore, the first part of the Routh–
Hurwitz criterion for unconditional stability is satisfied. In order to check the second part of this criterion,
we consider a 2-dof model problem. The corresponding scalar form of (84) is
a3z3 þ a2z2 þ a1zþ a0 ¼ 0; ð85Þwhere
a3 ¼ a2 ¼ 1; a1 ¼Dt2
4x2
�þ e2
mu1
�þ Dt2
2x2
; a0 ¼
Dt2
4x2
�þ e2
mu
: ð86Þ
3222 J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232
Since Dt, u, x2, e2 and m > 0, all coefficients of the polynomial (85) are positive. Moreover
a1a2 � a0a3 ¼Dt4
8
e2x2
mu> 0; ð87Þ
which verifies that the second part of the Routh–Hurwitz criterion is satisfied and therefore the partitionedprocedure is unconditionally stable.
4.2.2.2. Consistency and accuracy. Using the equilibrium equation at tn to eliminate €UUn in expressions (35),
(73) and (80) for bFF unþ1,
bFF/nþ1 and
_UUnþ1, respectively, and substituting the resulting effective load vectors into
(78) and (79), then writing the resulting equations in a matrix form yield
dnþ1 ¼ Rdn þ Ln; ð88Þ
where dn is given in (63) and
R ¼ R�11 R2; Ln ¼ R�1
1 L0; ð89Þ
R1 ¼I þ Dt2M�1
uu Kuu=4 0 Dt2M�1uu Ku/=4
DtM�1uu Kuu=2 I DtM�1
uu Ku/=2
0 0 K// þ Dt2KTu/M
�1uu Ku/=4
264375;
R2 ¼I � Dt2M�1
uu Kuu=4 DtI �Dt2M�1uu Ku/=4
�DtM�1uu Kuu=2 I �DtM�1
uu Ku/=2
KTu/ � Dt2KT
u/M�1uu Kuu=2 DtKT
u/ �Dt2KTu/M
�1uu Ku/=4
264375;
L0 ¼Dt2M�1
uu ðFn þ Fnþ1Þ=4DtM�1
uu ðFn þ Fnþ1Þ=2Dt2KT
u/M�1uu ðFn þ Fnþ1Þ=4�Qnþ1
264375:
ð90Þ
Inserting the solution dðtnÞ into (88), we have
dðtnþ1Þ ¼ RdðtnÞ þ Ln þ sðtnÞ; ð91Þ
where sðtnÞ ¼ sTu ðtnÞ _ssTu ðtnÞ s/ðtnÞ� �T
is the local truncation error. Inserting the Taylor expansions (65)–
(67) into (91) and expanding the resulting equation yield
suðtnÞ � OðDt3Þ; _ssuðtnÞ � OðDt3Þ; s/ðtnÞ � OðDt3Þ; ð92Þ
sðtnÞ � OðDt3Þ and sðtnÞ ! 0 as Dt ! 0: ð93Þ
Therefore, the partitioned scheme is consistent and second-order accurate.
4.3. Explicit partitioned scheme
The explicit partitioned scheme can be also constructed by integrating the mechanical equation directly
using an explicit integrator, e.g. the central difference scheme for the velocity and acceleration fields
_UUn ¼1
2DtðUnþ1 �Un�1Þ; €UUn ¼
1
Dt2ðUnþ1 � 2Un þUn�1Þ: ð94Þ
J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232 3223
The mechanical equation of motion at tn is
Muu€UUn þ KuuUn þ Ku/Un ¼ Fn: ð95Þ
Substituting the second equation in (94) into (95) yields
1
Dt2MuuUnþ1 ¼ bPPn; ð96Þ
where
bPPn ¼ Fn þ2
Dt2Muu
�� Kuu
Un �
1
Dt2MuuUn�1 � Ku/Un: ð97Þ
The resulting algorithm for the explicit partitioned scheme is:
1. Evaluate the mechanical field by Eq. (96). Note that if lumped mass matrix is employed, there is no
solution of equations for the mechanical field.
2. Evaluate the electric field by
K//Unþ1 ¼ KTu/Unþ1 �Qnþ1: ð98Þ
3. If required, compute the velocity and acceleration fields at tn using Eq. (94).
Based on Eq. (94), the starting value of displacements is taken as
U�1 ¼ U0 � Dt _UU0 þDt2
2€UU0: ð99Þ
4.3.1. Stability of the explicit partitioned scheme
As before, we seek for a non-trivial solution in the form of (44). Substituting Eq. (97) into (96), and (44)
into (96) and (98) yields
�K// KTu/
Ku/ Kuu þ4z2
ð1� z2ÞDt2 Muu
24 35 Un
Un
� ¼ 0: ð100Þ
The characteristic polynomial of (100) is
det z2 Muu
��� Dt2
4Kuu
�þ Ku/K
�1//K
Tu/
�þ Dt2
4Kuu
�þ Ku/K
�1//K
Tu/
��¼ 0: ð101Þ
Since Kuu þ Ku/K�1//K
Tu/ is positive definite, by the theorem of Bellman, the stability condition of the
explicit partitioned scheme becomes
Muu �Dt2
4Kuu
�þ Ku/K
�1//K
Tu/
�P 0: ð102Þ
By similar arguments described in Section 4.1.1, it follows that the explicit partitioned scheme is con-
ditionally stable and the condition for stability is governed by
Dt6Dtexcr ¼ 2=xexn ; ð103Þ
where xexn is the largest value of the following generalized eigenvalue problem
Kuu
�hþ Ku/K
�1//K
Tu/
�� ðxexÞ2Muu
ix ¼ 0: ð104Þ
3224 J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232
Since Kuu is positive definite, by comparing Eqs. (53) and (104), we have
xexn > xim
n ; and thus Dtexcr < Dtimcr : ð105Þ
4.3.2. Consistency and accuracy
Substituting Eq. (97) into (96) and (98), we have
ynþ1 ¼ A1yn þ A2yn�1 þ Ln; ð106Þwhere
yn ¼Un
Un
� ; A1 ¼
2I � Dt2M�1uu Kuu �Dt2M�1
uu Ku/
K�1//K
Tu/ 2I � Dt2M�1
uu Kuu
��Dt2K�1
//KTu/M
�1uu Ku/
� ; ð107Þ
A2 ¼�I 0
�K�1//K
Tu/ 0
� ; Ln ¼
Dt2M�1uu Fn
Dt2K�1//K
Tu/M
�1uu Fn � K�1
//Qnþ1
� : ð108Þ
Inserting the exact solution yðtnÞ into (106) yields
yðtnþ1Þ ¼ A1yðtnÞ þ A2yðtn�1Þ þ Ln þ rðtnÞ; ð109Þwhere rðtnÞ ¼ rTu ðtnÞ rT/ðtnÞ
� �Tis the local truncation error. Inserting the Taylor expansions
yðtnþ1Þ ¼ yðtnÞ þ Dt _yyðtnÞ þDt2
2€yyðtnÞ þ
Dt3
6vyðtnÞ þOðDt4Þ; ð110Þ
yðtn�1Þ ¼ yðtnÞ � Dt _yyðtnÞ þDt2
2€yyðtnÞ �
Dt3
6vyðtnÞ þOðDt4Þ; ð111Þ
and (67) into (109), yield
ruðtnÞ � OðDt4Þ; r/ðtnÞ � OðDt4Þ: ð112ÞSubstituting the exact solutions into (94), we have
_UUðtnÞ ¼1
2DtUðtnþ1Þ½ �Uðtn�1Þ� þ _rruðtnÞ; ð113Þ
€UUðtnÞ ¼1
Dt2Uðtnþ1Þ½ � 2UðtnÞ þUðtn�1Þ� þ €rruðtnÞ; ð114Þ
where _rruðtnÞ and €rruðtnÞ are local truncation errors in velocity and acceleration, respectively. Substituting the
expansions (110) and (111) into (113) and (114), yield
_rruðtnÞ � OðDt2Þ; €rruðtnÞ � OðDt2Þ: ð115ÞThe truncation errors vanish as Dt ! 0. Therefore, the explicit partitioned scheme is consistent.
5. The RVE-based multilevel method for partitioned scheme
The explicit and implicit conditionally stable as well as implicit unconditionally stable partitionedschemes proposed in the previous section entail solving linear systems of equations at each time station.
Attention is restricted to piezocomposites where the size of the microstructure is comparable to that of the
wavelength of a traveling signal necessitating discretization on the scale of heterogeneity. The classical
multigrid approach with standard linear interpolation operators applied to the resulting discrete system has
J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232 3225
been show to exhibit poor rate of convergence, mainly because the lower frequency subspace is not geo-metrically smooth for problems in heterogeneous media [6].
We focus on the RVE-based multilevel method developed by the authors for structural dynamics
problems [6], in which the special intergrid transfer operators are constructed from the solution of the RVE.
The so called RVE-based multilevel approach has been shown to possess an increasing rate of convergence
with increase in mismatch of properties between microconstituents and decrease in time step size. In the
case of homogeneous media and static analysis, the method has been shown to recover the rate of con-
vergence of the classic multigrid method, i.e., 1/3.
We note that in the mechanical field equations (38) and (79), the coefficient matrix is identical to thatobtained from the conventional structural dynamics problems. Convergence characteristics of the RVE-
based multilevel method for mechanical equations has been studied in [6]. In the remaining of this section
attention is restricted on convergence studies of the RVE-based multilevel method applied to the electric
field equations (38) or (98), and (78).
5.1. Two-level iteration matrices
We consider the two-level iteration process consisting of t SSOR pre- and post-smoothing iterations.The coarse-model dielectric and effective dielectric stiffness matrices are obtained by restricting the corre-
sponding source grid matrices:
K0// ¼ QTK//Q; bKK 0// ¼ eQQT bKK //eQQ; ð116Þ
where Q and eQQ are the prolongation operators for the corresponding electric equations. The resulting
two-level iteration matrices are given by
L// ¼ St//T//S
t//;
bLL// ¼ bSS t//
bTT//bSS t
//; ð117Þ
where
S// ¼ I � P�1//K//; bSS// ¼ I � bPP�1
//bKK //; ð118Þ
T// ¼ I �QK�10//Q
TK//; bTT// ¼ I � eQQ bKK�10//
eQQT bKK //; ð119Þ
P// and bPP// are smoothing pre-conditioners.
5.2. Prolongation operators
5.2.1. One-dimensional case
The RVE is composed of three elements as shown in Fig. 1. Between adjacent RVEs a ‘‘soft’’ (withsmaller stiffness) interface element is placed. For simplicity, we focus on the case of constant mass density,
q1 ¼ q2 ¼ q, and the volume fraction a ¼ 0:5.The effective dielectric and dielectric stiffness matrices of the RVE are evaluated as
bKK r// ¼ K r
// þDt2
4K rT
u/ðM ruuÞ
�1K r
u/; ð120Þ
K r// ¼
a1 �a1 0 0
�a1 a1 þ b1 �b1 0
0 �b1 a1 þ b1 �a10 0 �a1 a1
26643775; a1 ¼ e1A=le; b1 ¼ e2A=le; ð121Þ
RVE
le le
c1 e1 ε1, , c2 e2 ε2, , ρ1 ρ2=
Fig. 1. The 1D model with RVE.
3226 J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232
K ru/ ¼
a2 �a2 0 0
�a2 a2 þ b2 �b2 0
0 �b2 a2 þ b2 �a20 0 �a2 a2
26643775; a2 ¼ e1A=le; b2 ¼ e2A=le; ð122Þ
M ruu ¼ qleAdiagð1=2; 1; 1; 1=2Þ; ð123Þ
where bKK r//, K
r// and K r
u/ are the effective dielectric, dielectric and piezoelectric stiffness matrices of the
RVE, respectively; M ruu is the lumped mass matrix of the RVE; q, A and le are the mass density, cross-
sectional area and element length, respectively; e1, e2, e1 and e2 are dielectric and piezoelectric constants of
the two material constituents, respectively. The prolongation operators for the RVE are constructed basedon the corresponding constrained minimization problems:
Minimize : P1r ¼ 1
2wT
r Kr//wr subjected to kwrk2 ¼ 1; ð124Þ
Minimize : P2r ¼ 1
2/T
rbKK r
///r subjected to k/rk2 ¼ 1 ð125Þ
which yields the following eigenvalue problems:
K r//wr ¼ lrwr; kwrk2 ¼ 1; ð126Þ
bKK r///r ¼ kr/r; k/rk2 ¼ 1: ð127Þ
The global prolongation operators Q and eQQ are block diagonal matrices:
Qn�m ¼ diagfQr;Qr; . . . ;Qrg; eQQn�m ¼ diagf eQQr;eQQr; . . . ;
eQQrg; ð128Þ
Qr ¼ ½ p1 p2 �; eQQr ¼ ½ q1 q2 �; ð129Þwhere p1 and p2 are eigenvectors corresponding to the first two smallest eigenvalues of the eigenvalue
problem (126), whereas q1 and q2 are eigenvectors corresponding to the first two smallest eigenvalues of the
eigenvalue problem (127).
5.2.2. Two-dimensional case
We consider a 2D structured mesh in x1 and x3 plane as shown in Fig. 2. Plane stress assumption is made.
The RVE consists 9 elements as illustrated in Fig. 2. The total electric field degrees of freedom for the RVE
problem is 16.
As in the 1D case, the global prolongation operators Q and eQQ are assembled from the prolongation
matrices, Qr andeQQr, of RVEs. The RVE prolongation matrix consists of eigenvectors of the eigenvalue
The 9-element RVE
(1) (2)
Two phases of the RVE
Fig. 2. Global and RVE (local) finite element meshes.
J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232 3227
problems (126) and (127), respectively. In the numerical examples considered, we take the first 8 eigen-
vectors corresponding to the lowest eigenvalues.
The two constituent phases are assumed to be made of PZT ceramic with transversely isotropic pro-
perties. The constitutive equation (3) in matrix form can be written as
r ¼ CS � eTE; D ¼ eS þ eE; ð130Þwhere
r ¼ ½ r1 r2 r3 r4 r5 r6 �T; S ¼ ½ S1 S2 S3 S4 S5 S6 �T; ð131Þ
D ¼ ½D1 D2 D3 �T; E ¼ ½E1 E2 E3 �T; ð132Þ
Ckl ¼
c11 c12 c13 0 0 0
c12 c11 c13 0 0 0
c13 c13 c33 0 0 0
0 0 0 c44 0 0
0 0 0 0 c44 0
0 0 0 0 0 ðc11 � c12Þ=2
2666666664
3777777775; eij ¼
e11 0 0
0 e11 0
0 0 e33
264375;
eik ¼0 0 0 0 e15 0
0 0 0 e15 0 0
e31 e31 e33 0 0 0
264375; k; l ¼ 1; 2; . . . ; 6; i; j ¼ 1; 2; 3: ð133Þ
Tensor and matrix indices are related by
11 ! 1; 22 ! 2; 33 ! 3; 23 ! 4; 31 ! 5; 12 ! 6
or alternatively
k ¼ 12ðiþ jÞdij þ ½9� ðiþ jÞ�ð1� dijÞ;
3228 J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232
where dij is the Kronecker delta. Components of stress and strain tensors and matrices are related as
Stress: rij ! rk;
Strain: Sij ! Sk for i ¼ j (k ¼ 1; 2; 3); 2Sij ! Sk for i 6¼ j (k ¼ 4; 5; 6).
For plane stress in x1 and x3 plane r2 ¼ r4 ¼ r6 ¼ 0 and E2 ¼ 0.
The constitutive equation (130) in the matrix form can be simplified as
r1
r3
r5
24 35 ¼~cc11 ~cc13 0~cc13 ~cc33 00 0 c44
24 35 S1S3S5
24 35�0 ~ee310 ~ee33e15 0
24 35 E1
E3
� ; ð134Þ
D1
D3
� ¼ 0 0 e15
~ee31 ~ee33 0
� S1S3S5
24 35þ e11 0
0 ~ee33
� E1
E3
� ; ð135Þ
where
~cc11 ¼ c11 � c212=c11; ~cc13 ¼ c13 � c12c13=c11; ~cc33 ¼ c33 � c213=c11;
~ee31 ¼ e31 � e31c12=c11; ~ee33 ¼ e33 � e31c13=c11; ~ee33 ¼ e33 þ e231=c11:ð136Þ
The spectral radii of the two-level iteration matrices qðL//Þ and qðbLL//Þ with t ¼ 3 are evaluated nu-
merically and the results are presented in Tables 1–4, in which n is the size of the source mesh andDtcr ¼ 2=xim
n is the critical time step for the implicit conditionally stable partitioned scheme; ~ee11 ¼ e11,~cc44 ¼ c44 and ~ee15 ¼ e15.
For the explicit and implicit conditionally stable partitioned schemes, the coefficient matrix of the electric
field equation is simply the dielectric stiffness matrix. It only depends on the dielectric constants of the
media. From Tables 1 and 2, it can be observed that the rate of convergence improves with increasing
heterogeneity in dielectric constants and is independent of the problem size.
The coefficient matrix of the unconditionally stable partitioned scheme is the effective dielectric
matrix, which depends on the time step, the dielectric and piezoelectric constants. From Tables 3 and 4,
Table 1
Spectral radius of the 1D two-level iteration matrix qðL//ÞqðS3
//T//S3//Þ n ¼ 100 n ¼ 200 n ¼ 300
e1=e2 ¼ 1 0.0243 0.0243 0.0243
e1=e2 ¼ 50 0.0206 0.0206 0.0206
e1=e2 ¼ 100 0.0116 0.0116 0.0116
e1=e2 ¼ 1000 0.0013 0.0013 0.0013
Table 2
Spectral radius of the 2D two-level iteration matrix qðL//ÞqðS3
//T//S3//Þ n ¼ 14� 14 n ¼ 22� 22 n ¼ 30� 30
~eeð1Þii =~eeð2Þii ¼ 1 0.0216 0.0356 0.0420
~eeð1Þii =~eeð2Þii ¼ 50 0.0204 0.0212 0.0214
~eeð1Þii =~eeð2Þii ¼ 100 0.0136 0.0138 0.0139
~eeð1Þii =~eeð2Þii ¼ 1000 0.0020 0.0021 0.0021
Table 4
Spectral radius of the 2D two-level iteration matrix qðbLL//Þ
q bSS 3//
bTT//bSS 3
//
� �n ¼ 22� 22 n ¼ 30� 30
Dt=Dtcr ¼ 1 Dt=Dtcr ¼ 10 Dt=Dtcr ¼ 1 Dt=Dtcr ¼ 10
r1 ¼ 1, r2 ¼ 1 0.0726 0.4414 0.0808 0.4901
r1 ¼ 1, r2 ¼ 100 0.0360 0.2121 0.0366 0.2254
r1 ¼ 1, r2 ¼ 1000 0.0055 0.1322 0.0056 0.1386
r1 ¼ 10, r2 ¼ 1 0.0659 0.5116 0.0749 0.5621
r1 ¼ 10, r2 ¼ 100 0.0963 0.7231 0.0983 0.7872
r1 ¼ 10, r2 ¼ 1000 0.0249 0.6088 0.0252 0.6598
r1 ¼ 100, r2 ¼ 1 0.1073 0.0386 0.1290 0.0416
r1 ¼ 100, r2 ¼ 100 0.0187 0.9997 0.0195 0.9999
r1 ¼ 100, r2 ¼ 1000 0.0386 0.9999 0.0402 1.0000
r1 ¼ ~eeð1Þik =~eeð2Þik , r2 ¼ ~eeð1Þii =~ee
ð2Þii .
Table 3
Spectral radius of the 1D two-level iteration matrix qðbLL//Þ
q bSS 3//
bTT//bSS 3
//
� �n ¼ 100 n ¼ 200
Dt=Dtcr ¼ 1 Dt=Dtcr ¼ 10 Dt=Dtcr ¼ 1 Dt=Dtcr ¼ 10
e1=e2 ¼ 1, e1=e2 ¼ 1 0.0767 0.8558 0.0767 0.8569
e1=e2 ¼ 1, e1=e2 ¼ 100 0.0301 0.2846 0.0302 0.2846
e1=e2 ¼ 1, e1=e2 ¼ 1000 0.0039 0.1170 0.0039 0.1170
e1=e2 ¼ 4, e1=e2 ¼ 1 0.0783 0.6484 0.0784 0.6485
e1=e2 ¼ 4, e1=e2 ¼ 100 0.0638 0.8378 0.0639 0.8380
e1=e2 ¼ 4, e1=e2 ¼ 1000 0.0102 0.4262 0.0102 0.4262
e1=e2 ¼ 10, e1=e2 ¼ 1 0.9976 0.9999 0.9994 1.0000
e1=e2 ¼ 10, e1=e2 ¼ 100 0.0871 0.9185 0.0871 0.9187
J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232 3229
we can observe that the rate of convergence improves with increasing heterogeneity in dielectric con-stants, but increasing heterogeneity in piezoelectric constants tends to affect the rate of convergence
adversely. In all cases, the rate of convergence improves with decreasing time step and is independent of
the problem size.
To this end we compare the critical time steps of the implicit conditionally stable and the explicit par-
titioned schemes in 2D case. Numerical results reveal that the ratio of critical time steps in implicit and
explicit schemes is only slightly affected by the heterogeneity in dielectric and piezoelectric constants, but is
extremely sensitive to the heterogeneity in elastic constants. In Fig. 3, we plot the ratio of critical time steps
between the implicit and explicit schemes as well as the spectral radius of the two-level iteration matrixfor the mechanical equation versus the ratio of elastic constants ~ccð1Þkl =~cc
ð2Þkl .
It can be observed from Fig. 3 that with increasing ratio of elastic constants of the two constituent
phases, the spectral radius of the two-level iteration matrix for the mechanical equation decreases and the
ratio of critical time steps increases. Thus the implicit conditionally stable partitioned scheme can use a
larger time step than that used for the explicit scheme at the expense of solving a linear system of mecha-
nical equations at each time station. When the mismatch of mechanical properties is considerable, the
100 200 300 400 500 600 700 800 900 10006
8
10
12
14
16
18
20
22
24
Rat
io o
f crit
ical
tim
e st
eps
Ratio of critical time steps ∆tim/∆tex
Spectral radius for the
mechanical equation
100 200 300 400 500 600 700 800 900 10000.005
0.0094
0.0139
0.0183
0.0228
0.0272
0.0317
0.0361
0.0406
0.045
Ratio of elastic constants c(1)/c(2)
Spe
ctra
l rad
ius
of it
erat
ion
mat
rix fo
r m
echa
nica
l equ
atio
n
Fig. 3. Ratio of critical time steps and spectral radius of mechanical equation versus the ratio of elastic constants.
3230 J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232
spectral radius of the two-level iteration matrix for the mechanical equation could be sufficiently small for
the two-level iteration process to converge in a couple of cycles.
5.3. Convergence of the two-level method
Consider the problem illustrated in Fig. 1. The bar is fixed at (x ¼ 0) and free at (x ¼ 1) with no externalload and electric charge acting on the bar. The following initial disturbance in the displacement field is
considered:
f ðxÞ ¼ f0a0½x� ðx0 � dÞ�4½x� ðx0 þ dÞ�4f1� H ½x� ðx0 þ dÞ�g½1� Hðx0 � d� xÞ�;where a0 ¼ 1=d8 and HðxÞ is the Heaviside step function; f0, x0 and d are the magnitude, the location of the
maximum value and the half width of the initial pulse. This pulse is similar in shape to the Gaussian
distribution function. In the computations, we take x0 ¼ 1=2 and f0 ¼ 0:05.For both electric and mechanical equations, we employ three SSOR pre- and post-smoothing iterations.
The stopping criterion is taken as
krk2kf k2
6Tol ¼ 10�6; ð137Þ
where krk2 and kf k2 are the 2-norms of the residual and the right-hand-side vectors, respectively.
For parametric study of convergence characteristics of the RVE-based multilevel solver, we carry out
computations with different ratios of material constants. The results are summarized in Tables 5 and 6 for
the implicit conditionally and unconditionally stable partitioned schemes, respectively. In both tables, the
numbers in the parentheses correspond to the cycle count of the mechanical equation, while the numbers
without parentheses denote the cycle numbers for the electric field equation. For the conditionally stablescheme, we take Dt ¼ Dtimcr .
From Tables 5 and 6, it can be seen that the RVE-based multilevel method converges faster for the
mechanical equation than for the electric equation. Moreover, increasing the ratios of elastic and dielectric
constants, and cutting the time step size tend to reduce the number of cycles.
Table 6
Numbers of cycles for the unconditionally stable scheme (Tol ¼ 10�6)
Iteration numbers n ¼ 400 n ¼ 1000
Dt=Dtcr ¼ 1 Dt=Dtcr ¼ 10 Dt=Dtcr ¼ 1 Dt=Dtcr ¼ 10
c1=c2 ¼ 1, e1=e2 ¼ 1, e1=e2 ¼ 1 6(2) 64(4) 5(2) 87(4)
c1=c2 ¼ 1, e1=e2 ¼ 1, e1=e2 ¼ 10 5(2) 26(4) 5(2) 29(4)
c1=c2 ¼ 1, e1=e2 ¼ 1, e1=e2 ¼ 100 4(2) 9(4) 4(2) 9(4)
c1=c2 ¼ 1, e1=e2 ¼ 1, e1=e2 ¼ 1000 3(2) 6(4) 3(2) 6(4)
c1=c2 ¼ 1, e1=e2 ¼ 4, e1=e2 ¼ 100 5(2) 55(4) 5(2) 65(4)
c1=c2 ¼ 10, e1=e2 ¼ 2, e1=e2 ¼ 2 4(4) 37(4) 4(3) 37(4)
c1=c2 ¼ 100, e1=e2 ¼ 2, e1=e2 ¼ 2 4(4) 37(3) 4(3) 36(3)
c1=c2 ¼ 1000, e1=e2 ¼ 2, e1=e2 ¼ 2 4(3) 35(2) 4(2) 35(2)
Table 5
Numbers of cycles for the conditionally stable scheme (Tol ¼ 10�6)
Iteration numbers n ¼ 400 n ¼ 1000 n ¼ 2000
c1=c2 ¼ 1, e1=e2 ¼ 1, e1=e2 ¼ 1 4(2) 4(2) 4(2)
c1=c2 ¼ 1, e1=e2 ¼ 1, e1=e2 ¼ 10 4(2) 4(2) 4(2)
c1=c2 ¼ 1, e1=e2 ¼ 1, e1=e2 ¼ 100 3(2) 3(2) 3(2)
c1=c2 ¼ 1, e1=e2 ¼ 1, e1=e2 ¼ 1000 2(2) 2(2) 2(2)
c1=c2 ¼ 1, e1=e2 ¼ 4, e1=e2 ¼ 100 3(2) 3(2) 3(2)
c1=c2 ¼ 10, e1=e2 ¼ 2, e1=e2 ¼ 2 4(4) 3(3) 3(3)
c1=c2 ¼ 100, e1=e2 ¼ 2, e1=e2 ¼ 2 4(3) 3(3) 3(3)
c1=c2 ¼ 1000, e1=e2 ¼ 2, e1=e2 ¼ 2 4(2) 3(2) 3(2)
J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232 3231
Acknowledgements
The support of the Sandia National Laboratories under Contract DE-AL04-94AL8500 and the Office
of Naval Research through grant number N00014-97-1-0687 are gratefully acknowledged.
References
[1] C.A. Felippa, K.C. Park, C. Farhat, Partitioned analysis of coupled mechanical systems, Comput. Methods Appl. Mech. Engrg.
190 (2001) 3247–3270.
[2] S. Piperno, C. Farhat, Partitioned procedures for the transient solution of coupled aeroelastic problems, Part II: energy transfer
analysis and three-dimensional applications, Comput. Methods Appl. Mech. Engrg. 190 (2001) 3147–3170.
[3] E. Tursa, B.A. Schrefler, On consistency, stability and convergence of staggered solution procedures, Rend. Mat. Acc. Lincei,
Rome 5 (9) (1994) 265–271.
[4] J. Fish, V. Belsky, Multigrid method for a periodic heterogeneous medium. Part I: Convergence studies for one-dimensional case,
Comput. Methods Appl. Mech. Engrg. 126 (1995) 1–16.
[5] J. Fish, V. Belsky, Multigrid method for a periodic heterogeneous medium. Part 2: Multiscale modeling and quality control in
multidimensional case, Comput. Methods Appl. Mech. Engrg. 126 (1995) 17–38.
[6] J. Fish, W. Chen, RVE based multilevel method for periodic heterogeneous media with strong scale mixing, Journal of
Engineering Mathematics 46 (1) (2003) 87–106.
[7] H.F. Tiersten, Linear Piezoelectric Plate Vibrations, Elements of the Linear Theory of Piezoelectricity and the Vibrations of
Piezoelectric Plates, Plenum Press, New York, 1967.
[8] Y.H. Lim, V.V. Varadan, V.K. Varadan, Finite-element modeling of the transient response of MEMS sensors, Smart Mater.
Struct. 6 (1997) 53–61.
[9] C.Q. Chen, X.M. Wang, Y.P. Shen, Finite element approach of vibration control using self-sensing piezoelectric actuators,
Comput. Struct. 60 (3) (1996) 505–512.
3232 J. Fish, W. Chen / Comput. Methods Appl. Mech. Engrg. 192 (2003) 3211–3232
[10] Y. Kagawa, T. Tsuchiya, T. Kataoka, Finite element simulation of dynamic responses of piezoelectric actuators, J. Sound Vib. 191
(4) (1996) 519–538.
[11] D.A. Saravanos, P.R. Heyliger, D.A. Hopkins, Layerwise mechanics and finite element for the dynamic analysis of piezoelectric
composite plates, Int. J. Solids Struct. 34 (3) (1997) 359–378.
[12] O.J. Aldraihem, R.C. Wetherhold, Mechanics and control of coupled bending and twisting vibration of laminated beams, Smart
Mater. Struct. 6 (1997) 123–133.
[13] T.J.R. Hughes, The Finite Element Method, Linear Static and Dynamic Finite Element Analysis, Prentice-Hall, Englewood Cliffs,
NJ, 1987.
[14] T.J.R. Hughes, Analysis of transient algorithms with particular reference to stability behavior, in: T. Belytschko, T.J.R. Hughes
(Eds.), Computational Methods For Transient Analysis, North-Holland, Amsterdam, New York, Oxford, 1983.
[15] C. Farhat, K.C. Park, Y.D. Pelerin, An unconditionally stable staggered algorithm for transient finite element analysis of coupled
thermoelastic problems, Comput. Methods Appl. Mech. Engrg. 85 (1991) 349–365.
[16] O.C. Zienkiewicz, D.K. Paul, A.H.C. Chan, Unconditionally stable staggered solution procedure for soil-pore fluid interaction
problems, Int. J. Numer. Meth. Engrg. 26 (1988) 1039–1055.
[17] J.D. Lambert, Computational Methods in Ordinary Differential Equations, Wiley, London, 1973.
[18] R. Bellman, Introduction to Matrix Analysis, McGraw-Hill, New York, 1960.