Journal of Computational Physics 229 (2010) 6193–6207

Contents lists available at ScienceDirect

Journal of Computational Physics

journal homepage: www.elsevier.com/locate/jcp

Finite element modeling of the deformation of magnetoelastic film q

Matthew I. Barham a, Daniel A. White a,*, David J. Steigmann b
a
Lawrence Livermore National Laboratory, 7000 East Avenue, Livermore, CA 94551, United States
b
6133 Etcheverry Hall, University of California, Berkeley, CA 94720, United States

a r t i c l e i n f o a b s t r a c t

Article history: Recently a new class of biocompatible elastic polymers loaded with small ferrous particles,
Received 28 May 2009 a magnetoelastic polymer, has been developed. This engineered material is formed into a
Received in revised form 2 April 2010 thin film using spin casting. An applied magnetic field will deform the film. The magnetic
Accepted 7 April 2010
deformation of this film has many possible applications, particularly in microfluidic pumps
Available online 29 April 2010
and pressure regulators. In this paper a finite element method suitable for the transient
simulation of arbitrarily shaped three-dimensional magnetoelastic polymers subjected to
Keywords:
time-varying magnetic fields is developed. The approach is similar to that employed in
Magnetoelastic
Magnetics
finite elment magnetohydrodynamic simulations, the key difference is a more complex
Deformation hyperelastic material model. In order to confirm the validity of the approach, finite element
Finite element solutions for an axially symmetric thin film are compared to an analytical solution based
on the membrane (infinitely thin) approximation. For this particular problem the two
approaches give qualitatively similar results and converge as the film thickness approaches
zero.
Published by Elsevier Inc.

1. Introduction

The motivation for this work stems from the creation of thin magnetoelastic films. The magnetoelastic material consists
of a mixture of iron powder, polydimethylsiloxane (PDMS), and surfactants. The surfactant is used to increase dispersion of
the iron particles. Excellent mixing is a requirement to optimize the dispersion of the iron into the PDMS matrix and to break
up large iron agglomerates. The material is formed into a thin film using spin casting. Some possible applications of magneto-
elastic films include microfluidic actuators [1] and environmental sensors [2–4].
The basic equilibrium theory for magnetoelastic materials can be found in, for example, [5–7]. In these works the energy,
stress, and variational principles are derived. There are differences depending upon the choice of primary fields, but for par-
ticular choices the net result is that the total system stress tensor can be written in terms of the elastic Cauchy stress tensor
[8] that is independent of the magnetic field, plus a Maxwell stress tensor [9] that is independent of displacement. Since fi-
nite element methods for elasticity and magnetics exist, the total system stress tensor concept points the way to a fully-cou-
pled magnetoelastic finite element method.
The goal of the work presented here is to develop a computational procedure for the simulation of arbitrarily shaped
three-dimensional magnetoelastic films subjected to a transient magnetic field. An unstructured mesh is used to represent
the geometry, a finite element method is used for the spatial discretization, and implicit finite difference methods are used
for the temporal discretization. The elastic equations are solved for the material deformation, the magnetic equations are

q
This work was performed under the auspices of the US Department of Energy by the University of California, Lawrence Livermore National Laboratory
under contract DE-AC52-07NA27344.
* Corresponding author. Tel.: +1 925 422 9870.
E-mail address: white37@llnl.gov (D.A. White).

0021-9991/$ - see front matter Published by Elsevier Inc.

doi:10.1016/j.jcp.2010.04.007
6194 M.I. Barham et al. / Journal of Computational Physics 229 (2010) 6193–6207

solved for the magnetic fields in both the material region and the large but finite volume of surrounding space. The surround-
ing space contains coils used to generate the time-dependent magnetic field. The magnetic equations are not magnetostatic,
but rather a time-dependent diffusion equation, i.e. eddy current equation. This is important if the time dependence of the
external field is fast enough (e.g. kilohertz) to generate eddy currents. The time-dependent diffusion equation developed
here does converge to the magnetostatic solution for steady-state problems. The elastic equations and the magnetic equa-
tions are solved in an operator split manner. This approach has successfully been used in magnetohydrodynamic (MHD) sim-
ulations [10]. Some key differences for this magnetoelastic application, compared to the MHD application, are that a
hyperelastic Mooney–Rivlin model is used for the PDMS material constitutive relation, and the time scale warrants an
unconditionally stable implicit time integration method.
In order to confirm the validity of the approach, finite element solutions are compared to an analytical solution for the
special case of a thin axially symmetric magnetoelastic film in a static dipole field. The analytical method is based on the
membrane approximation, the aspect ratio between the thickness and radius of the film is small, thus taking the leading order
terms of an asymptotic expansion leads to a model where rigidity with respect to bending is neglected [11]. In the limit as
the film thickness approaches zero and the deformation approaches steady-state, the finite element solution to approaches
the analytical solution, providing an important verification of the finite element implementation.
In Section 2 the finite element formulations is discussed including subsections on the governing partial differential equa-
tions, variational statement, finite element basis functions, time integration, mesh relaxation and solution process. Section 3
gives an outline of the analytic model used for comparison. Section 4 presents the values used for the calculation and the
results. A conclusion is provided in Section 5.

2. Finite element formulation

2.1. Governing partial differential equations

We begin by briefly reviewing various formulations for time-dependent low-frequency electromagnetics, i.e. the diffusion
or eddy current approximation. Maxwell’s equations involve the electric ð~ ~ fields, the electric ð~
EÞ and magnetic ðHÞ DÞ and
magnetic ð~BÞ flux densities, and the current density ~ J. These quantities are not independent but are related to each other
through the constitutive relations ~D ¼ ~
E; ~
B ¼ lH,~ and ~J ¼ r~
E where  is the dielectric permittivity, l is the magnetic per-
meability, and r is the electrical conductivity. Since there are divergence constraints
~ ~
r J ¼ 0; ð1Þ
~ ~
r B ¼ 0; ð2Þ
it is sometimes convenient to introduce the magnetic ð~
AÞ and electric ð~
TÞ vector potentials such that
~ ~ ~
B¼r A; ð3Þ
~ ~ ~
J¼r T; ð4Þ
and in order to uniquely define ~ A and ~
T gauge conditions need to be applied to these vector potentials. Depending upon the
chosen gauge conditions additional electric (UE) and magnetic (UH) scalar potentials may be introduced. The redundancy of
the fields and arbitrariness of gauge conditions leads to a great variety of finite element formulations for electromagnetics.
An example of an ~ E-field formulation can be found in [12], in this formulation the electric field is the only variable. Examples
~
of the H-field formulation can be found in [13,14], in this formulation the magnetic field is the only variable. Examples using
the magnetic vector potential and the electric scalar potential, the so-called ~ A  U formulation, can be found in [15–17]. It
has been shown in [18] that, when using H(Curl)-conforming basis functions, these three formulations give identical results,
the only difference is that the different methods require different boundary to be provided. Yet another formulation uses the
electric vector potential and the magnetic scalar potential, the so-called ~ T  X formulation, as described in [19]. This ap-
proach supposedly has advantages for problems in which only a small portion of the entire problem is conducting.
The approach used here is to work with the primary fields rather than with vector potentials. A combined ~ E ~ B formu-
lation is used, i.e. the electric field and the magnetic flux density are both computed. This approach is used because a mag-
netic field is required for computation of magnetic forces, and the electric field facilitates coupling with external circuits [20].
And a pragmatic reason is that we are building upon an existing magnetohydrodynamic code [10] that uses ~ E and ~B as the
variables. The key electromagnetic partial differential equations are then
~  rr
r ~ UE ¼ 0; ð5Þ

r~ ~  1~
Eind ¼ r ~ UE ;
B þ rr ð6Þ
l
@~
B ~ ~
¼ r Eind ; ð7Þ
@t
1 1 ~ ~
TM ¼ ~ B ~
B B  B I; ð8Þ
l 2l
 M.I. Barham et al. / Journal of Computational Physics 229 (2010) 6193–6207 6195

where the first equation is a Poisson’s equation for the electrostatic potential, the second equation is Ampere’s law
(neglecting displacement current, i.e. the eddy current approximation), and the third equation is Faraday’s law. The elec-
trostatic potential (UE) is used to drive current into a coil, if a coil is present in the problem. In (6) ~
Eind is the induced, or
solenoidal, component of the electric field. As mentioned previously there are also divergence constraints on the electro-
magnetic quantities (1) and (2). These divergence constraints can be satisfied by proper choice of finite element basis
functions.
The above electromagnetic equations are coupled with the Cauchy equation of motion via the total stress tensor T,
2
d~ u~  T þ q~
q 2
¼r b; ð9Þ
dt
T ¼ TC þ TM: ð10Þ
In (9) q is the density, ~ u is the displacement, T is the total stress tensor, and ~b is a body force (e.g. gravity). Note that (9) is
derived by combining conservation of linear momentum with conservation of mass. Conservation of angular momentum is
also satisfied by (9) when the stress tensor is symmetric. For the magnetoelastic film application the Cauchy stress tensor
will be given by the nonlinear hyperelastic Mooney–Rivlin model. In this model the stress T C is a function of the displace-
ment ~ u. The functional form is complicated and requires some further definitions.
Following [8], we define a set of points in R3 describing the body of interest as the configuration M of the body. Points in
M are denoted as ~ X ¼ fX 1 ; X 2 ; X 3 g 2 M and are called material points, these labels move with the material. Points in R3 are
denoted as ~ x ¼ fx1 ; x2 ; x3 g 2 R3 are called the spatial points. The coordinate system for R3 is referred to as the spatial system
or laboratory frame, while the coordinate system for M is referred to as the material frame. There exists a smooth, invertible,
time-dependent mapping ~ x ¼ nð~X; tÞ; n : M ! R3 that describes the motion of M.
It is convention that nð~ X; 0Þ ¼ ~X, the two coordinate frames coincide at time t = 0. Given two nearby points ~ X and ~X þ d~X,
at some later time t we have
@~
x ~
xð~
~ xð~
X; tÞ  ~ X þ d~
X; tÞ  dX ¼ F d~
X; ð11Þ
~
@X
where F is the deformation gradient, with FiA = @xi/@XA. In general the deformation consists of a composition of rotation and
stretch, this can be expressed as F ¼ VR where R is an orthonormal rotation matrix and V 2 ¼ B ¼ FF T is a symmetric positive
definite tensor called the left Cauchy–Green deformation tensor. The tensor V is called the left stretch tensor. Since V is sym-
metric positive definite it has the decomposition V ¼ M K  M T where M is a unitary matrix defining an orthogonal coordinate
system and K is a diagonal matrix with three entries k1, k2, k3, the eigenvalues of V. The values k1, k2, k3 are called the prin-
ciple stretches and correspond to the amount of stretching along the three orthogonal coordinate axes. The Jacobian, the ra-
tio between deformed volume and reference volume, is given by
qffiffiffiffiffiffiffiffiffiffiffiffiffiffi
J¼ detðBÞ ¼ detðFÞ ¼ k1 k2 k3 : ð12Þ

An alternative decomposition of the deformation gradient is F ¼ RU where R is again an orthonormal rotation matrix and
U 2 ¼ C ¼ F T F is a symmetric positive definite tensor called the right Cauchy–Green deformation tensor and U is called the
right stretch tensor. The eigenvalues of V equal the eigenvalues of U. In the former case of d~ x ¼ VRd~X the line segment
d~ x ¼ RUd~
X is rotated then stretched, in the latter case of d~ X the line segment is stretched then rotated.
The displacement of a material point is defined as
   
u ~
~ x ~
X; t ¼ ~ X; t  ~
X: ð13Þ

The displacement gradient is

@~
u
¼ F  I: ð14Þ
@~
X
There are numerous strain measures, the Lagrangian finite strain tensor is a measure of how much the right Cauchy–Green
deformation tensor differs from I,
1
E¼ ðC  IÞ: ð15Þ
2
For the PDMS film of interest here, there is significant deformation and the linear elasticity approximation is not adequate. A
hyperelastic material is a material in which the Cauchy stress is function of a scalar strain energy density function W, where
W is a function of the invariants of the deformation. Here we will employ an isotropic, homogeneous, hyperelastic model
called the Mooney–Rivlin constitutive relation. The strain energy density is given in terms of the invariants of the Cau-
chy–Green deformation tensor,
!
1
W ¼ aðI1  3Þ þ bðI2  3Þ þ c  1 þ dðI3  1Þ2 ; ð16Þ
I23
6196 M.I. Barham et al. / Journal of Computational Physics 229 (2010) 6193–6207

where c ¼ 12 a þ b; d ¼ að5m2Þþbð11
2ð12mÞ
m5Þ
, a and b are independent phenomenological constants related to the shear moduli
through G = 2(a + b), m is Poisson’s ratio, {I1, I2, I3} are the invariants of the Cauchy–Green deformation tensor and can be rep-
resented in terms of the principal stretches (k1, k2, k3) as: I1 ¼ k21 þ k22 þ k23 ; I2 ¼ k21 k22 þ k22 k23 þ k21 k23 and I3 ¼ k21 k22 k23 . The Cauchy
stress T c can be expressed in terms of the Piola–Kirchoff stress S, and the Piola–Kirchoff stress can be expressed in terms of
the strain energy W,

1
Tc ¼ FSF T ;
J
1
E ¼ ðF T F  IÞ;
2
@W
S¼ :
@E
A straightforward but tedious calculation results in the formula for the Cauchy stress tensor as a function of the deformation
" !#
1 2 1
Tc ¼ 2ða þ bI1 ÞB  2bB þ 4 c 2 þ dI3 ðI3  1Þ I : ð17Þ
J I3

2.2. Variational statements

The variational statements are derived by multiplying the partial differential equations above by suitable testing func-
tions, integrating over the domain, and invoking integration by parts formulas. Due to different compatibility requirements,
different testing functions will be used for the different partial differential equations.
For Cauchy’s equation of motion we require ~ u 2 ðHðGradÞÞ3 , the displacement is a fully continuous vector field. The var-
iational form of (9) is constructed by multiplying by a test vector U~ 2 ðHðGradÞÞ3 and integrating over the entire domain X,
Z Z Z
@ 2~
u ~
q  U dv ¼ ~  TÞ  U
ðr ~ dv þ q~ ~ dv :
bU ð18Þ
X @t2 X X
~ is considered to have units of distance, then each term in the above equation has units of work, and hence
If the test vector U
this variational method is often referred to as the Method of Virtual Work. Integration by parts is employed to yield
Z Z   Z Z
@ 2~
u ~ ~ dv þ q~
~ U ~ dv þ ðT n ~ da;
q  U dv ¼ T: r bU ^Þ  U ð19Þ
X @t2 X X C

where the notation A : B represents the tensor dot product, e.g. A : B ¼ Aij Bij , the symbol  denotes the vector outer product,
and C is the boundary of the domain X with an outward unit normal of n ^.
In Poisson’s equation we require UE 2 H(Grad), the potential is a continuous scalar field. The variational form of (5) is stan-
dard, the equation is multiplied by a scalar test function X 2 H(Grad) and integrated over the domain X to yield
Z
~  rr
ðr ~ UE ÞX dv ¼ 0:
X

Green’s first scalar identity is used to obtain

Z I
~ UE  r
rr ~ X dv ¼ ~ UE X da;
^  rr
n ð20Þ
X C

which is to be satisfied for all test functions X.

Ampere’s law requires ~ Eind 2 HðCurlÞ, the electric field must have a finite curl. Across material interfaces the tangential
component of the ~ Eind is continuous, but the normal component may be discontinuous. The variational form is obtained
by multiplying Ampere’s law (6) by a test function W ~ 2 HðCurlÞ and integrating over the three-dimensional problem domain
X to obtain the variational form
Z Z Z
r~ ~ dv ¼
Eind  W ~  1~
r ~ dv þ
BW ~ UE  W
rr ~ dv : ð21Þ
X X l X

Now integration by parts is used on (21) and with the Gauss divergence theorem the result is
Z Z Z I
1~ ~ ~ ~ UE  W 1
r~ ~ dv ¼ 
Eind  W B  r  W dv þ rr ~ dv þ ^ ~
n ~ da:
BW ð22Þ
X X l X C l
The resulting surface integral term has units of electrical current and will become important when we discuss boundary con-
ditions; but for the sake of clarity, we will omit this term for the remaining derivation.
The magnetic flux density must satisfy ~ B 2 HðDiv Þ; ~
B must have a finite divergence, and across material interfaces the
normal component must be continuous, but the tangential component may be discontinuous. The variational form of
 M.I. Barham et al. / Journal of Computational Physics 229 (2010) 6193–6207 6197

Faraday’s law is obtained by multiplying by a test function ~

V 2 HðDiv Þ and integrating over the three-dimensional problem
domain X to yield
Z Z
@~
B ~ ~ ~
 V dv ¼  r Eind  ~
V dv : ð23Þ
X @t X

No integration by parts is required for this equation.

The natural and essential boundary conditions for (19) are

^ ¼ 0;
Natural T n
ð24Þ
~
Essential u ¼ 0:
In general, the inhomogeneous boundary conditions require vector-valued functions ~
t b and ~
ub such that

^ ¼~
Tn tb on Ct ;
~
u¼~ ub on Cu ;
where Cu and Ct are the portion of the boundary where displacement and traction are prescribed respectively, note that
Cu [ Ct = C .
The natural and essential boundary conditions for (20) are
~ UE ¼ 0;
^  rr
Natural n
ð25Þ
Essential UE ¼ 0:
In other words, the normal component of the conduction current density rr ~ UE is the natural boundary condition while the
surface scalar potential (or voltage) UE is the essential boundary condition. In general, the inhomogeneous versions of these
two boundary conditions require scalar valued functions gN and gD such that

n ~ UE ¼ g
^  rr on CN ;
N
UE ¼ g D on CD ;
where CN and CD are portion of the boundary where current density and scalar potential are prescribed respectively note
that CN [ CD = C.
The natural and essential boundary conditions for (22) are
1
^ ~
Natural n B ¼ 0;
l ð26Þ
^ ~
Essential n Eind ¼ 0:
In general, the inhomogeneous versions of these two boundary conditions require vector-valued functions ~
g N and ~
g D such
that
1
^ ~
n B ¼~
gN on CN ;
l
^ ~
n E ¼~
gD on CD :

2.3. Finite element basis functions

The variation forms above employ four different function spaces, U ~  ðHðGradÞÞ3 for the displacement, X  H(Grad) for the
~ ~
scalar potential, W  HðCurlÞ for the electric field, and V  HðDiv Þ for the magnetic flux density. We will use different finite
element basis functions for each of these spaces. The basis functions are defined on a computational mesh, for our applica-
tion the mesh is a conforming unstructured mesh of hexahedral elements. Each basis function is constructed to preserve the
characteristic of the quantities they are used to represent.
The simplest basis function that is X  H(Grad) conforming is a collection of piecewise linear nodal interpolatory basis
functions. This can be represented by WI for I = 1 . . . N for each mesh node I where N is the total number of nodes. This
can be interpolated on nodes meaning that if ~ xJ is the position of node J where J = 1 . . . N, then

WI ð~
xJ Þ ¼ dIJ ; ð27Þ
The electric scalar potential UE at a position of ~
x is then given by the basis function expansion
X
N
UE ð~
xÞ ¼ /I WI ð~
xÞ; ð28Þ
I¼1

where /I is the value of the potential at mesh node I.

6198 M.I. Barham et al. / Journal of Computational Physics 229 (2010) 6193–6207

The functions U ~  ðHðGradÞÞ3 are discretized using the basis functions WI for each component of the vector. This can be
represented as W~ k where W
~1 ¼ ^ ~2 ¼ ^
x1 WI ; W ~3 ¼ ^
x2 WI , and W x3 WI are vector-valued basis functions in the x1, x2, and x3 direc-
I I I I
tions. This can be interpolated on nodes meaning that if ~ xi xiJ where i = 1 . . . 3 is the position of node J = 1 . . . N, then
xJ ¼ ^
~ i xj ¼ dJK dij ;
W ð29Þ
I J

where j = 1 . . . 3. The basis function expansion of the displacement ~

u at a position of ~
x is given by

X
N
~
uð~
xÞ ¼ ~ i ð~
uiI W I xÞ ð30Þ
I¼1

respectively. The coefficient uiI is the value of the ith component of the displacement at mesh node I. It is more notationally
convenient to eliminate the superscript i and instead allow the index K to run from K = 1 . . . 3N, with
uK ¼ uiI ; i ¼ K mod 3; I ¼ ðK  1Þ=3 þ 1.
There has been much analysis of various basis functions for the spaces H(Curl) and H(Div). We do not provide any original
analysis here, rather we will simply review the key properties of the so-called ‘‘edge” H(Curl)-conforming basis functions and
‘‘face” H(Div)-conforming basis functions. The edge basis functions W ~  ðHðCurlÞÞ are piecewise linear basis functions de-
fined on the computational mesh. This basis functions can be represented as H ~ A for A = 1 . . . M, where M is the number of
mesh edges, these functions interpolate on edges, meaning that if ^t B is the edge vector of edge B = 1 . . . M, then
Z
~ A  ^t B dl ¼ dAB ;
H ð31Þ
L

where the line integral is over mesh edge A. The electric field at position ~
x is given by the basis function expansion
X
M
~
Einc ð~
xÞ ¼ ~ A ð~
eA H xÞ; ð32Þ
A¼1

where eA is the induced voltage along edge A. It is convenient to implement the edge basis functions on a reference hexahe-
dron and then transform to the actual element. Eq. (31) implies that the appropriate transformation for the basis functions is

~ A ¼ F T H
H ~0 ; ð33Þ
A

where F is the deformation tensor defined in the discussion of the Mooney–Rivlin material model, and prime denotes the
reference basis function. This transformation preserves the continuity of the tangential component of the field across ele-
ment material interfaces, but allows for the normal component to be discontinuous, which is the correct physical continuity
of electric fields.
The face basis functions ~V  ðHðDiv ÞÞ are also piecewise linear basis functions defined on the computational mesh. This
basis functions can be represented as K ~C for C = 1 . . . F, where F is the number of mesh faces, interpolate on faces, these basis
functions meaning that if n^ D is an face normal vector D = 1 . . . F, then
Z
~C  n
K ^ D da ¼ dCD ; ð34Þ
A

where the surface integral is over mesh face C. The magnetic flux density at position ~
x is given by the basis function
expansion
X
F
~
Bð~
xÞ ¼ ~C ð~
bC K xÞ; ð35Þ
D¼1

where bC is the net flux through face C. It is convenient to implement the face basis functions on a reference hexahedron and
then transform to the actual element. Eq. (34) implies that the appropriate transformation for the basis functions is

~C ¼ 1 F T K
K ~0 ; ð36Þ
C
J
where F is the deformation tensor, J is the element Jacobian, and prime denotes the reference basis function. This transfor-
mation preserves the continuity of the normal component of the field across element material interfaces, but allows for the
tangential component to be discontinuous, which is the correct physical continuity of magnetic flux.
Another key property of the node, edge and face basis functions is the inclusion relations,

~ X 2 W;
r ~ ð37Þ
~ W
r ~ 2~
V ð38Þ
which is simple version of the De-Rham complex [21]. These inclusion relations allow for discrete versions of the vector
~ r
identities r ~ f ¼ 0 and r
~ r~  g ¼ 0 to be satisfied exactly.
 M.I. Barham et al. / Journal of Computational Physics 229 (2010) 6193–6207 6199

Using the basis functions described above the discrete variation forms can be written as

@2
M u þ KðuÞ ¼ f; ð39Þ
@t 2
S/ ¼ g; ð40Þ
MW e ¼ ðDWV ÞT b þ DXW /  r; ð41Þ
@
MV b ¼ DWV e; ð42Þ
@t
where u, /, e, and b are the arrays of coefficients for the basis function expansion of the displacement, the electrostatic po-
tential, the electric field, and the magnetic field, respectively. The matrices are given by
Z
MKL ¼ qW ~ L dX;
~K  W ð43Þ
X
Z
K KL ðuÞ ¼ ~W
T K ðuÞ : r ~ L dX; ð44Þ
X
Z
SIJ ¼ rr~ WI  r~ WJ dX; ð45Þ
X
Z
MW AB ¼ qH~A  H~ B dX; ð46Þ
ZX
MVCD ¼ K~C  K
~D dX; ð47Þ
ZX
1~ ~ ~
DWV
AC ¼ r  HA  KC dX; ð48Þ
X l
Z
DXW
IA ¼
~ WI  H
r ~ A dX: ð49Þ
X

The various basis functions, quadratures, and mappings for all of the above matrices have been implemented in the FEMSTER
library [22,23]. The matrices are computed using standard finite element procedures, i.e. each subscript pair (KL, IJ, AB, CD, AC
and IA) interaction is computed in a reference element using Gaussian quadrature, this result is transformed to the actual
element, and the final result is accumulated in a sparse matrix data structure. Note that in the material stiffness matrix K
the stress tensor is a complicated nonlinear function of both the displacement u and the magnetic fields, and for simplicity
the stress is assumed constant over each element. Also note that Eq. (42) can be simplified using the inclusion relations, the
matrix DWV can be written as DWV = MVKWV, where KWV is an edge-face incidence matrix. Therefore (42) can be written as
@
b ¼ KWV e; ð50Þ
@t
it is an explicit relation. As another side note, the divergence constraints (1) and (2) can be expressed as

ðDXW ÞT e ¼ 0; ð51Þ
WV
K b ¼ 0: ð52Þ
XW XW W XW XW
The matrix D can be decomposed as D = M K , where K is a node-edge incidence matrix. From the topological iden-
tity KWVKXW = 0, it is clear that the discrete divergence conditions (51) and (52) are satisfied automatically, for any mesh,
without the need to introduce Lagrange multipliers or a penalty method. In words, at every node in the mesh the edge-based
electric fields sum to zero, and at every element in the mesh all face-based magnetic fluxes sum to zero. This is one advan-
tage of using the edge and face basis functions for finite element electromagnetics.
The arrays f, g, and r are the coefficients for the independent source terms. The array f involves a body force such as grav-
ity and a surface traction, the array g involves current injected into the problem domain, and the array r involves a surface
current density,
Z Z
fK ¼ q~ ~ K dX þ
bW ðT n ~ K dC;
^Þ  W ð53Þ
C
ZX
gI ¼ ðn ~ UÞWI da;
^  rr ð54Þ
C
Z
1
rA ¼ ^ ~
n ~ A da:
BH ð55Þ
C l

2.4. Time integration

The elastic equations and the Maxwell equations will be integrated in time implicitly, but in an operator split manner.
When updating the displacement the electromagnetic fields are held constant, and when updating the electromagnetic fields
6200 M.I. Barham et al. / Journal of Computational Physics 229 (2010) 6193–6207

the displacement is held constant. Operator splitting is a well known approximation often used in multi-physics simulations,
the error can be shown to be of first order accurate with respect to the time step Dt.
In the semi-discrete equation of motion (39) the displacement coefficients u are continuous functions of time. Time inte-
gration is achieved by using finite difference methods in which u is known at discrete instants. Let superscript n denote the
2
discrete-time index, with tn+1  tn = Dt. Let a and v denote @t@ 2 u and @t@ u, respectively. The Newmark family of time integration
methods is given by

Manþ1 þ Kðunþ1 Þ ¼ Fnþ1 ; ð56Þ

vnþ1 ¼ v n þ Dt½ð1  cÞan þ canþ1 ; ð57Þ
  
1
unþ1 ¼ u n þ Dt v n þ Dt 2  b an þ banþ1 ; ð58Þ
2
where c and b are arbitrary parameters in the intervals [0, 1] and 0; 12 , respectively. For linear elasticity it is common to solve
(56) for the acceleration an+1, this acceleration is used in (57) to update the velocity vn+1, and the velocity is used in (58) to
update the displacement un+1. For the linear case the accuracy and stability of the method can be analyzed. The special values
of b = 1/4 and c = 1/2 give the trapezoidal rule, or average acceleration method, which is second-order accurate and uncon-
ditionally stable. However, since the stress T is a nonlinear function of u for the Mooney–Rivlin model, it is necessary to ex-
press an+1 in terms of un+1,
1 1 1 n 1  2b n
anþ1 ¼ unþ1  2 un  v  a ð59Þ
Dt 2 b Dt b Dtb 2b
and then this is used in (56) to give
1 1 1 n 1  2b n
Munþ1 þ Kðunþ1 Þ ¼ Fnþ1 þ M 2 un þ M v þM a ð60Þ
Dt 2 b Dt b Dtb 2b
giving a nonlinear equation for un+1 in terms of known quantities.
For the magnetic film application it is desirable to add a controlled amount of dissipation into the time integration algo-
rithm in order to damp out high-frequency waves and to more rapidly converge to a steady-state solution. The Hilber–
Hughes–Taylor integrator modifies (56) by adding an additional parameter a,

Manþ1 þ ð1 þ aÞKðunþ1 Þ ¼ aKðun Þ þ Fnþ1 ; ð61Þ

where a is an independent parameter. This results in the final form of the full-discrete equation of motion,
1 1 1 n 1  2b n
Munþ1 þ ð1 þ aÞKðunþ1 Þ ¼ aKðun Þ þ Fnþ1 þ M 2 un þ M v þM a : ð62Þ
Dt 2 b Dt b Db 2b
The same Eqs. (57) and (58) are used to integrate acceleration and velocity. The HHT integration method will be uncondi-
tionally stable and second-order accurate provided a 2 [1/3, 0] and
1  2a
c¼ ; ð63Þ
2
ð1  aÞ2
b¼ : ð64Þ
4
To form a discrete-time set of equations for the electromagnetic coefficients e, b and / are known at discrete time inter-
vals denoted by the subscript integer n. The fully discrete form of Poisson’s equation and Ampere’s law become

SX /nþ1 ¼ gnþ1 ; ð65Þ

nþ1
MW enþ1 ¼ ðDWV ÞT b þ DXW /nþ1  rnþ1 : ð66Þ
A generalized trapezoidal approximation is used for the time derivative of the magnetic field such that
n nþ1
nþ1 n @b @b
b ¼ b þ ð1  xÞDt þ xDt : ð67Þ
@t @t
The averaging parameter x determines the nature of the numerical time integration such that
8
< 0;
> Explicit; 1st Order Accurate Forward Euler;
x ¼ 1=2; Implicit; 2nd Order Accurate Crank Nicolson;
>
:
1; Implicit; 1st Order Accurate Backward Euler:
Applying this discretization to the semi-discrete Faraday’s law (50), gives
nþ1 n
b ¼ b  Dtðð1  xÞKWV en þ xKWV enþ1 Þ: ð68Þ
 M.I. Barham et al. / Journal of Computational Physics 229 (2010) 6193–6207 6201

~ defined as
An intermediate variable is introduced, b,
~ n ¼ bn  ð1  xÞDtKWV en :
b ð69Þ
n+1
Combining (69), (68), and (66) yields the final update equation for e ,
~ n þ DXW /nþ1 :
ðMW þ xDtKWV Þenþ1 ¼ ðDWV ÞT b ð70Þ

2.5. Mesh relaxation

The above finite element equations for the electromagnetics are valid in the material frame, meaning the computational
mesh is fixed to the material and moves with the material. This is ideal for coupling with the equations of finite element
elasticity since these equations are also cast in the material frame. However a key difference between the electromagnetics
and the elasticity is that magnetic fields exist in the vacuum surrounding the magnetic film, and a region of vacuum sur-
rounding the film must be included in the calculation. The computational mesh is fixed to the film and hence moves with
the film, however if the mesh in the vacuum region does not somehow also move, the mesh can become severely distorted
and eventually become unacceptable. To remedy this the computational mesh of the vacuum region is allowed to relax.
Numerous algorithms exist for mesh relaxation, the method used here is equipotential relaxation which essentially moves
each mesh node to the mean value of all surrounding nodes, this can be shown to be a solution to Poisson’s equation for the
node coordinates. The important concept is that, as the mesh in the vacuum region is adjusted, the magnetic field must be
interpolated from the old mesh to the new mesh in a manner that preserves the properties of the field. This interpolation is
often referred to as advection, because it is identical to the advection operator r  ~ v  ~B that appears in the Eulerian solution
of magnetic problems with motion. However in the case of mesh relaxation, the velocity ~ v ¼ D~x=Dt is not a physical velocity,
rather it is just the velocity of the mesh as it is relaxing. It is known that naive numerical methods for the magnetic advection
operator result in non-physical oscillations of the fields (reminiscent of Gibb’s or Runge’s phenomena) hence sophisticated
upwind methods with flux limiting must be used. In addition, the advection is complicated by the fact that the magnetic field
must remain divergence free. A detailed derivation of a divergence-preserving flux-limited advection algorithm is presented
in [10].

2.6. Complete finite element solution process

Here we provide a high-level summary of the finite element solution process for the fully-coupled magnetoelastic equa-
tions. The solution sequence is

Step 1. Compute the Maxwell stress tensor (8) using magnetic fields at time tn.
Step 2. Update the displacement un+1 using Eq. (62). During this step the Maxwell stress is constant, but the Cauchy stress is
given by the hyperelastic Mooney–Rivlin model (17), hence (62) is a nonlinear system of equations. This system of
equations is solved using Newtons’s method, with a multigrid preconditioner [24,25].
Step 3. Update the velocity (57) and the acceleration (58) and store these values for the next displacement calculation.
Step 4. Update Poisson’s equation (65) for the electrostatic potential. The boundary conditions on UE are chosen to drive a
current into a coil, generating a dipole-like external magnetic field. This is a linear implicit equation and is s solved
using a the same multigrid preconditioner as in Step 2.
Step 5. Update the electric field using (66). This is a linear implicit equation and is solved using a different multigrid pre-
conditioner specifically developed for edge-based discretization of curl–curl equations [26].
Step 6. Update the magnetic field using (50), this is an explicit equation.
Step 7. In the vacuum region, if there has been significant motion of the mesh, relax the mesh and interpolate the magnetic
coefficients bn+1 onto the new mesh. No other electromagnetic variables needs to be interpolated as they will be
computed ‘‘from scratch” during the next update.

This completes one cycle of the magnetoelastic solution process. The process is repeated using a constant time step Dt.
The time integration is unconditionally stable, hence the time step is chosen based on accuracy requirements.

3. Analytical model

The above finite element formulation of the magnetoelastic equations is quite general, it can be applied to arbitrary three-
dimensional geometries, the applied magnetic field is arbitrary, and the approach can be used to investigate fast transients,
or sinusoidal response, or steady-state response. In order to test the finite element formulation it is necessary to construct a
simple problem that has an analytical solution. In this section we review the analytical solution to a thin circular magneto-
elastic membrane in an external magnetic field. The derivation of this solution can be found [11].
The starting point is the equation of magnetoelastic equilibrium. The geometry is symmetric and the solution is one-
dimensional, i.e. the membrane height as a function of radius. It is assumed that the magnetization of the membrane itself
6202 M.I. Barham et al. / Journal of Computational Physics 229 (2010) 6193–6207

is weak, i.e. the self-field is neglected. The membrane approximation is used, the aspect ratio between film thickness and film
radius is small and only the leading oder term in the asymptotic expansion is used. In the membrane approximation there is
no resistance to bending. The Mooney–Rivlin model for hyperelasticity is used for the stress as a function of displacement.
And finally the external magnetic field is given by an ideal dipole field. This leads to a set of coupled differential equations
that is solved numerically using a shooting method,
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
k ¼ u=r; j¼ u0 2 þ z0 2 ; cos h ¼ u0 =j;
" " # !#
H ðh  zÞ3 ðh  zÞ2
~ jj Þ1 w
j0 ¼ ðrw ~ kj ðj cos h  u=rÞ þ r
~j  w
~ k cos h  w 6
4 6
sin h þ 1 þ 4 6
u cos h ;
l l l
" " # !#
3 2
0 1 H ðh  zÞ ð h  zÞ
~ jÞ
h ¼ ðr w ~ k sin h þ r
w 6
4 6
cos h þ 1 þ 4 6
u sin h ;
l l l
where r 2 [0, r0] is the radius from the axis of symmetry of a material point in the reference configuration; k(r) and j(r) are
the azimuthal and radial principal stretches; u(r) and z(r) are the radial and axial coordinates of a material point in the de-
formed configuration; /(r) is an angle defined in Fig. 1; h is the assigned distance between the dipole and the mid-plane of
the undeformed film; l is the distance from the dipole to a deformed material point (l2 = u2 + (h  z)2; w
~ ¼ tf W; H = 3D2l0vtf;
W is the conventional strain energy function dependent on the principal stretches; D is the dipole strength; l0 is the free-
space permeability; v is the films magnetic susceptibility; tf is the film thickness. Refer to Fig. 1 for a geometric represen-
tation. Note that the notation ()0 = d()/dr and Greek or Latin subscripts are used to denote partial derivatives.
We can solve for w ~ with the use of the Mooney–Rivlin strain energy function (Eq. (16)), incompressibility (k1k2k3 = 1) and
the relation between W and w ~ arriving at:

tf G
~ jÞ ¼
wðk; dðk2 þ j2 þ k2 j2  3Þ þ ð1  dÞðk2 þ j2 þ k2 j2  3Þ : ð71Þ
2
The above system is solved using a shooting method.
The goal is to compare full 3D finite thickness finite element simulations to this analytical model. It is seen that the only
place the film thickness appears in the analytical equilibrium equations is in H = 3D2l0vtf and G* = tfG. If the thickness is
parameterized by a constant e, as in tnew
f ¼ told
f , we can adjust the values of the dipole strength (D) and shear modulus (G)
as follows to maintain an equivalent comparison,
rffiffiffi
1
Dnew ¼ Dold ; ð72Þ
e
1
Gnew ¼ Gold : ð73Þ
e

4. Finite element results

The magnetoelastic film is a thin circular disk of thickness  and radius r0 = 0.317 cm is centered in the x  y plane, with
displacement in the z direction. The thickness will be varied using values tf = (140, 70, 35, 17.5) lm in order to investigate

Fig. 1. Geometry for the membrane approximation.

 M.I. Barham et al. / Journal of Computational Physics 229 (2010) 6193–6207 6203

deformation vs thickness. A small steel ring, with a fixed electric current, is used to create a magnetic field similar to a mag-
netic dipole. The ring is located a distance h above the membrane. The finite element formulation is inherently three-dimen-
sional, and hence it is possible to simulate deviations from symmetry, but to compare to the analytical membrane
approximation symmetry is enforced by appropriate boundary conditions. Only one quarter of the problem is meshed
around the axis of symmetry, i.e. the x P 0 and y P 0 quadrant, and displacement symmetry boundary conditions are im-
posed on the x = 0 and y = 0 planes. The geometry for the finite element simulation is illustrated in 1. For the thickness of
tf = (140, 70, 35) lm the film was four elements thick, for the thickness of tf = 17.5 lm the film was two elements thick.
The computational mesh consisted of 43,200 elements and 42,048 elements for these two cases. On all mesh boundaries that
are not symmetry planes the displacement is constrained to be zero. The displacement on the edge of the film is also con-
strained. The entire edge of the film is not constrained since this would impose zero slope as well as the desired zero dis-
placement. To get a better comparison to the membrane model, only the nodes on the mid-plane are constrained, this
allows for rotation at the edge of the film while maintaining zero displacement of the mid-plane at the edge. The geometry
is illustrated in Fig. 2.
Since the current carrying ring used in the finite element simulation does not produce an exact magnetic dipole field, an
equivalent dipole strength (Deq) for comparison to the analytical model is needed. Given the magnetic field produced by a
given current in the steel ring, an equivalent dipole strength is calculated for each mesh element between 0 6 r 6 r0 and
h + tf/2 6 z 6 7h/5 and the average of the equivalent dipole strengths will be used for comparison. The dipole magnetic field
used in the continuum model is:

3ða  kÞa  k
H¼D 3
; ð74Þ
l
where a is the unit vector from the dipole to a material point. The equivalent dipole strength is determined by taking the dot
product between the right hand side of (74) with the unit vector in z-direction and setting this equal to the z-component of
the magnetic field produced by the FEM resulting in:

2ðzc  hÞ2  r 2c
HzFEM ¼ Deq 5
; ð75Þ
l
and when the solving for the equivalent dipole strength:

5
l HzFEM
Deq ¼ ; ð76Þ
2ðzc  hÞ2  r2c
where zc is the height of the center of the element above the mid-pane of the film and rc is the radial position of the center of
the element away from the axis of symmetry. It should be noted that the relationship between the equivalent dipole strength
and the current density is linear. Thus once the equivalent dipole strength is calculated for one current density the equiv-

Fig. 2. Illustration of the geometry for the FEM simulations. This illustration is not to scale, the thickness of the membrane and the size of the current
carrying ring have been exaggerated for clarity.
6204 M.I. Barham et al. / Journal of Computational Physics 229 (2010) 6193–6207

alent dipole strength for any other current is known. For comparison of different finite element film thickness it convenient
to define Heq ¼ 3D2eq l0 vt f .
For comparison to the analytical membrane approximation, the finite element simulation is run to steady-state for a spe-
cific ring current, which corresponds to a specific effective dipole strength. Then the current is increased to the next effective
dipole strength, and the simulation is run to steady-state again. For each simulation, the time step was 10 ls. The number of
steps ranged from 15 to 100, more steps were required for the larger effective dipole strength as this gives larger displace-
ment. The process is illustrated in Fig. 3.

Fig. 3. Center displacement z0 for increasing dipole strength, increases of 1.553104 A m2 at each step.

h −5 h ALE, error = 3.3109% −5

r x 10 r x 10
5 5
0.15 4 0.15 4
0.1 3 0.1 3
z

2 2
0.05 1 0.05 1

0.1 0.2 0.3 0.1 0.2 0.3

r r
hz −5 hz ALE, error = 1.7241% −5
x 10 x 10

0.15 −4 0.15 −4
−6 −6
0.1 0.1
z

−8 −8
0.05 −10 0.05 −10
−12 −12
0.1 0.2 0.3 0.1 0.2 0.3
r r

hmag x 10
−5 hmag ALE, error = 1.6981% x 10
−5

12 12
0.15 10 0.15 10
0.1 8 0.1 8
z

6 6
0.05 4 0.05 4
2 2
0.1 0.2 0.3 0.1 0.2 0.3
r r

Fig. 4. Magnetic field model. The analytical analysis assumes an idealized infinitesimal magnetic dipole as the source of the magnetic field. The FEM
analysis employs a finite current ring. These plots show magnetic field magnitude, just above the membrane, for these two models.
 M.I. Barham et al. / Journal of Computational Physics 229 (2010) 6193–6207 6205

The following results are for a magnetoelastic film with the following properties and dimensions:

Properties Symbol Value

Radius of film r0 0.317 cm
Dipole height h 0.5 cm
(Film sheer modulus)  (film thickness) Gtf = G* 1750 N/m
Mooney–Rivlin tuning parameter d 0.9
Film magnetic susceptibility v 2.5
Film thickness tf 17.5  140 lm
Dipole strengths D 0  .0056 A2 m

Fig. 4 compares the magnetic field produced by the dipole to that of the finite steel ring used in the finite element model.
The error shown on the figure is the average error over the elements used to calculate the equivalent dipole strength. These
results show that the dipole approximation used in the analytical model is reasonable for this particular geometry, with a
maximum error of around 3.3%.
The results of the finite element simulation are shown in Fig. 5. These figures show the final deformation for given effec-
tive dipole strengths. A key observation is that the computational mesh in the air moves along with the membrane, this is the
mesh relaxation process described in Section 2.5 above. These FEM results are compared to the analytical results in Figs. 6
and 7. In Fig. 6 the displacement is shown versus position for a few discrete values of magnetic dipole strength. In Fig. 7 the
displacement at the membrane center is shown versus continuous magnetic dipole strength. The comparison is good,

Fig. 5. Displacement for the FEM model.

6206 M.I. Barham et al. / Journal of Computational Physics 229 (2010) 6193–6207

Fig. 6. Comparison of displacement, analysis vs FEM. In this graphic u is the position along the film radius, and z is the vertical displacement.

Fig. 7. Comparison of the z-displacement at the center of the film for the two methods as a function of dipole strength. This particular FEM result used a film
thickness of tf = 70 lm.

Fig. 8. This graphic compares the FEM results, with different film thicknesses, to the analytical approximation. The displacement is taken at the center of
the film. As the thickness decreases, the FEM results converge to the analytical membrane approximation.
 M.I. Barham et al. / Journal of Computational Physics 229 (2010) 6193–6207 6207

although not exact for numerous reasons. The results are not expected to match exactly because the analytical approxima-
tion employs the membrane approximation, i.e. no resistance to bending, and as discussed above the magnetic field created
by a small ring only approximately matches the dipole field used in the analytical approximation. Another reason for mis-
match is the finite film thickness used in the finite element simulations. In order to quantify this, several different film thick-
nesses were investigated. In Fig. 8 finite element results are compared to the analytical approximation for several different
film thicknesses. As the film thickness is decreased, the finite element results converge to the analytical approximation. This
provides a good verification of the finite element procedure. If the membrane model is assumed to be the exact solution the
error of the FEM model is: 37.0% for tf = 140 lm, 17.1% for tf = 70 lm, 9.7% for tf = 35 lm and 6.7% for tf = 17.5 lm.

5. Conclusions

A finite element algorithm for computing the magnetically driven displacement of a magnetoelastic film was developed.
This finite element model combines two established techniques, implicit solution of the Cauchy equation of motion using
H(Grad)3-conforming finite elements for the mechanics, and implicit solution of the low-frequency Maxwell’s equations
using a mixed H(Curl)–H(Div) formulation. The two finite element formulations are operator split in time, and coupled by
adding the Maxwell stress tensor to the Cauchy stress tensor for the total stress tensor. The algorithm is reminiscent of
an approach used for magnetohydrodynamics, the key difference is that the Mooney–Rivlin hyperelastic material model
is used for the material constitutive relation, hence at every time step a nonlinear system of equations must be solved.
The finite element algorithm is inherently three-dimensional and designed for transient simulations. However since there
is a lack of analytical test problems for full three-dimensional transient solution, the finite element algorithm was compared
to a previously developed analytical solution to a two-dimensional steady-state analysis of a circular magnetoelastic film
immersed in a dipole field. The analytical solution is based on the membrane approximation. By construction the finite ele-
ment algorithm must have a finite thickness film, and it is shown that as the film thickness is decreased the finite element
solution converges to the analytical membrane approximation, thus providing a good verification of the finite element
algorithm.

References

[1] E. Quandt, A. Ludwig, Magnetostrictive actuation in microsystems, Sensors Actuat. 81 (2000) 275–280.
[2] C. Grimes, K. Ong, K. Loiselle, P. Stoyanov, D. Kouzoudis, Y. Liu, C. Tong, F. Tefiku, Magnetoelastic sensors for remote query environmental monitoring,
Smart Mater. Struct. 8 (1999) 639–646.
[3] D. Kouzoudis, C. Grimes, The frequency response of magnetoelastic sensors to stress and atmospheric pressure, Smart Mater. Struct. 9 (2000) 885–889.
[4] D. Kouzoudis, C. Grimes, Remote query fluid flow velocity measurement using magnetoelastic thick film sensors, J. Appl. Phys. 87 (9) (2000) 6301–
6303.
[5] D. Steigman, Equilibrium theory for magnetic elastomers and magnetoelastic membranes, Int. J. Nonlinear Mech. 39 (2004) 1193–1216.
[6] S.V. Kankanala, N. Triantafyllidis, On finitely strained magnetorheological elastomers, J. Mech. Phys. Solids 52 (2004) 2869–2908.
[7] R. Bustamante, A. Dorfman, R.W. Ogden, On variational formulations in nonlinear magnetoelastics, Math. Mech. Solids 13 (2008) 725–745.
[8] J.E. Marsden, T.J.R. Hughes, Mathematical Foundations of Elasticity, second ed., Dover Publications, New York, 1983.
[9] J.D. Jackson, Classical Electrodynamics, second ed., Wiley, New York, 1975.
[10] R.N. Rieben, D.A. White, B.K. Wallin, J.M. Solberg, An arbitrary Lagrangian–Eulerain discretization of MHD on 3D unstructured grids, J. Comput. Phys.
226 (2007) 534–570.
[11] M. Barham, D.J. Steigman, M. McElfresh, R.E. Rudd, Finite deformation of a pressurized magnetoelastic membrane in a stationary dipole field, Acta
Mech. 191 (2007) 1–19.
[12] Z. Ren, F. Bouillault, A. Razek, A. Bossavit, J.C. Verite, A new hybrid model using electric field formulation for 3d eddy current problems, IEEE Trans. Mag.
26 (2) (1990) 470–473.
[13] A. Bossavit, J.C. Verite, The TRIFOU code: solving the 3d eddy current problems by using H as the state variable, IEEE Trans. Mag. 19 (6) (1983) 2465–
2470.
[14] A. Bossavit, Computational Electromagnetism: Variational Formulation, Complementarity, Edge Elements, Academic Press, New York, 1998.
[15] O. Biro, K. Preis, On the use of the magnetic vector potential in the finite element analysis of three-dimensional eddy currents, IEEE Trans. Mag. 25 (4)
(1989) 3145–3159.
[16] O. Biro, K. Preis, W. Renhart, K. Richter, G. Vrisk, Performance of different vector potential formulations in solving multiply connected eddy current
problems, IEEE Trans. Mag. 26 (2) (1990) 438–441.
[17] C. Bryant, C. Emson, C. Trowbridge, A comparison of Lorentz gauge formulations in eddy current problems, IEEE Trans. Mag. 26 (2) (1990) 430–433.
[18] R. Rieben, D. White, Verification of high-order mixed finite element solution of transient magnetic diffusion problems, IEEE Trans. Mag. 42 (1) (2006)
25–39.
[19] K. Fujiwara, Y. Okada, T. Nakata, N. Takahashi, Improvements in the t  x method for eddy current analysis, IEEE Trans. Mag. 24 (1988) 94–97.
[20] D.A. White, Using the Sherman–Morrison–Woodbury formula for coupling external circuits with FEM for eddy current problems, IEEE Trans. Mag. 45
(10) (2009) 3915–3918.
[21] D. Arnold, Compatible Spatial Discretization IMA. Volumes in Mathematics and its Applications, Springer, 2006.
[22] P. Castillo, J. Koining, R. Rieben, M. Stowell, D. White, Discrete differential forms: a novel methodology for robust computational electromagnetics,
Technical Report UCRL-ID-151522, Lawrence Livermore National Laboratory, Center for Applied Scientific Computing, January 2003.
[23] P. Castillo, R. Rieben, D. White, FEMSTER: An object oriented class library of higher-order discrete differential forms, ACM Trans. Math. Software 31 (4)
(2005) 425–457.
[24] M. Brezina, C. Tong, R. Becker, Parallel algebraic multigrid for structural mechanics, SIAM J. Sci. Comput. 27 (2006) 1727–1741.
[25] V. Hensen, U. Yang, BoomerAMG: a parallel algebraic multigrid solver and preconditioner, Appl. Numer. Math. 41 (2002) 155–177.
[26] Z. Kolev, P. Vassilevski, Parallel auxiliary space AMG for H(curl) problems, J. Comput. Math 27 (2009) 604–623.