You are on page 1of 6

IX CONGRESO COLOMBIANO DE MÉTODOS NUMÉRICOS: Simulación en Ciencias y Aplicaciones Industriales

IXCCMN 2013, Agosto. 21-23, 2013, UAO Cali, Colombia

A fully–discrete scheme for an eddy current problem with magnetically


nonlinear conductors
Ramiro Acevedo1 and Gerardo Loaiza1
1
Departamento de Matemáticas, Universidad del Cauca, Calle 5 # 4-70. Popayán, Colombia
e-mail: rmacevedo@unicauca.edu.co, gloaiza@unicauca.edu.co

Abstract
The aim of this work is to analyze a finite element fully-discrete approximation for a nonlinear magnetic field eddy current
problem in a bounded domain. The approximation is based in a formulation in terms of a vector magnetic potential and an electric
scalar potential [6, 7], where the Coulomb gauge condition is imposed by adding a penalization term, which is a suitable average
of the reluctivity in the computational domain. This work extends the analysis performed in [1] to the case of a magnetically
nonlinear conductor. We provide a backward-Euler fully-discrete approximation based on finite elements and show that the
resulting discrete variational problem is well posed by assuming physical properties of the magnetic reluctivity [5, 10, 11].
Furthermore, error estimates that prove optimal convergence are settled.
Key words: Nonlinear eddy current model, potential formulation, fully-discrete approximation, finite elements, error
estimates.

1 INTRODUCTION
The eddy current model is an approximation of the full Maxwell system of equations obtained by neglecting the displacement
current in Ampère’s law. This simplified model is suitable for most electrical engineering applications (the so-called low-
frequency regime), for instance, in the numerical simulation of electrical machines at power frequencies. Though the eddy
current equations are initially posed on the entire tree-dimensional space, we can switch to a bounded domain by introducing an
artificial boundary sufficiently removed from the region of interest.
The aim of this work is to analyze a finite element fully-discrete approximation for a nonlinear magnetic field eddy current
problem in a bounded domain. The approximation is based in a formulation in terms of a vector magnetic potential and an
electric scalar potential and extends the analysis performed in [1] to the case of a magnetically nonlinear conductor.
The outline of the paper is as follows: In Section 2, we introduce the nonlinear eddy current model in a bounded domain and
the potential-based formulation is deduced in Section 3. Next, in Section 4, we obtain a fully discrete scheme by approximating
the magnetic vector potential and scalar electric potential with nodal finite elements. Furthermore, the results presented in the
last two sections prove that the fully-discrete scheme is well-posed and convergent in an optimal way.
In what follows we will use standard notation for Sobolev spaces and norms.

2 A NONLINEAR MAGNETIC FIELD EDDY CURRENT PROBLEM


Let Ωc ⊆ R3 be the space occupied by the conductor material, Ω ⊆ R3 be a computational domain, which is an open and
bounded set such that Ωc ⊂ Ω and Ωd := Ω \ Ωc is the domain occupied by the dielectric material. The transient eddy current
model reads
curl H = J in Ω × [0, T ], (1)
∂B
+ curl E = 0 in Ωc × [0, T ], (2)
∂t
div B = 0 in Ω × [0, T ], (3)
where we have used standard notation in electromagnetism: H is the magnetic field, J the current density, B the magnetic
induction and E is the electric field. In order to obtain a closed system we need to add constitutive laws. Assuming that the
materials are electrically linear but magnetically nonlinear, we have
(
σE, in Ωc × [0, T ],
J= (4)
Jd , in Ωd × [0, T ],
and
H = H (B). (5)
Equation (4) is the Ohm’s law, where σ is the electrical conductivity of the conductor Ωc which is assumed to be a positive
and bounded function and Jd denotes an external density current source which is a data of the problem. We consider that Jd
is a free-divergence function and its support is included in the dielectric Ωd . In the magnetic constitutive relation (5), H is in
general a nonlinear mapping which can exhibit a history-dependent behavior (hysteresis) in the case of ferromagnetic materials.
But, in our analysis this nonlinear relationship will be represented through an anhysteretic H − B curve, which could have a
very steep slope (see [11]). More precisely, we consider (see, for instance [10] and [5])
(
νc (|B|)B in Ωc ,
H = ν(|B|)B := (6)
ν1 B in Ωd ,
where ν is the physical parameter called magnetic reluctivity (inverse of magnetic permeability), which will be considered a
positive constant in the dielectric. We assume that νc : R+ +
0 → R satisfies the following properties ([11, 5]):

• There exist positive constants νmin , νmax such that 0 < νmin ≤ νc (s) ≤ νmax for any s ≥ 0.
• The function s 7→ g(s) := νc (s)s is continuously differentiable and there exist positive constants κ0 and κ1 such that
0 < κ0 ≤ g 0 (s) ≤ κ1 ∀s ≥ 0. (7)

Finally, the complete eddy current model in a bounded domain is obtained by adding the following conditions about magnetic
induction field B:
B(x, 0) = B 0 (x) in Ω, (8)
B|Ωc × nc = B|Ωd × nc on ∂Ωc × [0, T ], (9)
B×n=0 on ∂Ω × [0, T ], (10)
where nc and n are the outward unit normal vectors to Ωc and Ω respectively. Let us remark that because of (3), the initial data
B 0 must be a free-divergence function in Ω. Furthermore, (9) in necessary to ensure that the curl of magnetic induction is a
square-integrable function in the whole Ω.

3 A POTENTIAL FORMULATION
We are going to start this section by recalling a (strong) classical formulation of the eddy current problem in terms of two
potentials: a magnetic vector potential A and an electric scalar potential v. We refer to Bı́ró & Preis [6] for a detailed discussion,
which also includes numerical tests showing the efficiency of this approach. Next, we will introduce a variational formulation
for the potential-based strong problem.
The starting point for the potential formulation is equation (3), which allows to look for a vector potential function A :
Ω × [0, T ] → R3 satisfying
B = curl A, div A = 0 in Ω × [0, T ], (11)
A×n=0 on Γ × [0, T ]. (12)
The free divergence condition for A is called the Coulomb gauge condition and together with the homogeneous boundary
condition (12) ensure the uniqueness for the vector potential A; see for instance [6] and [7]. Then, we can use equation (2) and
look for a scalar potential v : Ωc × [0, T ] → R3 satisfying
∂A ∂v
E= +∇ in Ωc × [0, T ]. (13)
∂t ∂t
We impose the gauge condition by adding a penalization term ν ∗ , which is a suitable average of ν in Ω (see [7]). Consequently,
the original eddy current equations in terms of potentials A and v can read as follows:

 
∂A ∂v
σ + σ∇ + curl (νc (curl A) curl A) + ν ∗ grad div A = 0 in Ωc × [0, T ], (14)
∂t ∂t
 
∂A ∂v
div σ + σ∇ =0 in Ωc × [0, T ], (15)
∂t ∂t
curl (ν1 curl A) + ν ∗ grad div A = Jd in Ωd × [0, T ], (16)
A·n=0 on Γ × [0, T ], (17)
ν1 curl A × n = 0 on Γ × [0, T ], (18)
 
∂A ∂v
σ + σ∇ · nc = 0 on Γc × [0, T ], p (19)
∂t ∂t
v(x, 0) = 0 a.e. x ∈ Ωc . (20)

(νc curl A) Ωc × nc = (ν1 curl A) Ω × nc on Γc × [0, T ]. (21)
d
A Ωc × nc = A Ω × nc on Γc × [0, T ]. (22)
d
A Ω · nc = A Ω · nc on Γc × [0, T ]. (23)
c d

We are now going to obtain a variational formulation for problem (14)–(20). First, we adapt the lines of [1, Section 4] to get
the following nonlinear parabolic weak formulation:
Find (A, v) ∈ L2 (0, T ; X × H1] (Ωc )) ∩ H1 (0, T ; L2 (Ωc )3 × H1] (Ωc )) such that

d
(A + ∇v, Z + ∇w)σ,Ωc + A (A, Z) = (Jd , Z)0,Ω ∀(Z, w) ∈ X × H1] (Ωc ),
dt (24)
A(·, 0) = A0 in Ω, v(·, 0) = 0 in Ωc ,

where the following notations have been used (Ω1c , . . . , Ωm


c are the connected components of Ωc ):
 Z 
X := H(curl; Ω) ∩ H0 (div; Ω), H1] (Ωc ) := u ∈ H1 (Ωc ) : u = 0, j = 1, . . . , m. ,
Ωjc
Z Z Z
A (A, Z) = νc (|curl A|) curl A · curl Z + ν1 curl A · curl Z + ν ∗ (div A)(div Z),
Ωc Ωd Ω
Z
(u, w)σ,Ωc := σu · w.
Ωc

The space X is endowed with the graph standard norm


1/2
kZkX := kZk20,Ω + k div Zk20,Ω + k curl Zk20,Ω .

Furthermore, as a consequence of the generalized Poincaré inequality (see [9, Lemma B63]), the seminorm |·|H1 (Ωc ) :=
k∇(·)k0,Ωc is a norm on H1] (Ωc ) which is equivalent to the H1 (Ωc )-norm.
Remark 1. To our knowledge, a proof of well-posedness for problem (24) has not been given yet. Actually, this proof is not
available even for the linear case (see [1]).

4 A FULLY DISCRETE SCHEME


In what follows we assume that Ω and Ωc are Lipschitz polyhedra (we recall that Ω is simply-connected). Let {Th }h be a regular
family of tetrahedral meshes of Ω such that each element K ∈ Th is contained either in Ωc or in Ωd . As usual, h stands for the
largest diameter of tetrahedra K in Th .
Consider the following finite element spaces:

X h := Z h ∈ X : Z h K ∈ P31 ∀K ∈ Th with K ⊂ Ω , Mh := uh ∈ H1] (Ωc ) : uh K ∈ P1 ∀K ∈ Th with K ⊂ Ω .


 

T
We consider a uniform partition {tn := n∆t : n = 0, . . . , N } of [0, T ] with a step size ∆t := N. For any finite sequence
{θn : n = 0, · · · , N }, let
n n−1
¯ n := θ − θ
∂θ , n = 1, 2, . . . , N.
∆t
The fully-discrete version of Problem (24) reads as follows:
Find (Anh , vhn ) ∈ X h × Mh , n = 1, 2, . . . , N such that:
¯ n + ∇∂v ¯ n , Z + ∇u + A (An , Z) = (Jd (tn ), Z)

∂A h h σ h 0,Ω ∀(Z, u) ∈ X h × Mh (25)
A0h = A0,h , vh0 = 0, (26)

where A0,h ∈ X h is a suitable approximation of A0 to obtain optimal error estimates.


In order to prove that Problem (25)–(26) has a unique solution, we first notice that at each iteration step we need to find
(Anh , vhn ) ∈ X h × Mh such that

(Anh + ∇vhn , Z + ∇u)σ + ∆t A (Anh , Z) = Fn (Z, u),


where

Fn (Z, u) = ∆t (Jd (tn ), Z)0,Ω + An−1 + ∇vhn−1 , Z + ∇u σ .



h

Consequently, to obtain the well-posedness of Problem (25)–(26), we first have to consider the nonlinear operator B :
H0 (curl; Ω) × H1] (Ωc ) → H0 (curl; Ω)0 × H1] (Ωc )0 given by

hB(A, v), (Z, u)i := (A + ∇v, Z + ∇u)σ,Ωc + ∆tA (A, Z)

and prove the properties of operator B given in the following lemma.

Lemma 1. The operator B is strongly monotone in X h × Mh , i.e., there exists a constant C > 0 such that
 
hB(A, v) − B(Z, u), (A, v) − (Z, u)i ≥ C kA − Zk2X + kv − uk2H1 (Ωc ) ∀(A, v), (Z, u) ∈ X h × Mh . (27)
]

Furthermore, B is Lipschitz continuous, i.e., there exists a constant C0 > 0 satisfying


 
kB(A, v) − B(Z, u)k ≤ C0 kA − Z)kX + ku − vkH1] (Ωc ) ∀(A, v), (Z, u) ∈ X h × Mh . (28)

Sketch of proof. We first recall that (7) implies the following inequalities for any α, β, η ∈ R3 (see, for instance, [4, Lemma 3]):

2
(νc (|α|)α − νc (|β|)β) · (α − β) ≥ κ
b0 |α − β| , (νc (|α|)α − νc (|β|)β) · η ≤ κ
b1 |α − β| |η| , (29)

for some positive constants κ


b0 and κ
b1 . Then, by using the first inequality in (29), we can easily check the following estimate for
any (A, v), (Z, u) ∈ X h × Mh :

hB(A, v) − B(Z, u), (A, v) − (Z, u)i


Z  Z Z 
2 2 2
≥ σmin |A − Z + ∇(v − u)| + ∆t C1 |curl(A − Z)| + ν ∗ |div(A − Z)| ,
Ωc Ω Ω

κ0 , ν1 }. Now, since Ω is a Lipschitz and simply-connected set, we use [3, Corollary 3.16] to get
where C1 := min{b

hB(A, v) − B(Z, u), (A, v) − (Z, u)i


Z Z 
2 2 2
≥ C2 |A − Z + ∇(v − u)| + |A − Z| + kA − ZkX .
Ωc Ω

Hence, by using the inequality


1 2 2
−2(A − Z) · ∇(v − u) ≤ 2 |A − Z| |∇(v − u)| ≤ |A − Z| + ε |∇(v − u)| ∀ε > 0,
ε
we conclude (27).
Finally, (28) is a direct consequence of the second inequality in (29).

Corollary 1. Problem (25)–(26) has a unique solution.

Proof. This result follows by combining the previous lemma and Zarantonello’s Theorem [12, Theorem 25.B].

5 ERROR ESTIMATES
In this section we will prove error estimates for our fully-discrete scheme. To this end, we have to assume that Problem (24) has
a unique solution and consider the projection operators Ph : X → X h and Qh : H1] (Ωc ) → Mh defined by

Ph Z ∈ X h : (Ph Z − Z, Y )X = 0 ∀Y ∈ X h

and

Qh u ∈ Mh : (∇Qh u − ∇u, ∇w)σ,Ωc = 0 ∀w ∈ Mh

It is a simple matter to see that lax-Milgram’s Lemma implies that Ph and Mh are well defined. Moreover, Cea’s Lemma yields
the following estimates.
Lemma 2. There exist positive constants C1 and C2 , independent of h:

kZ − Ph ZkX ≤ C1 inf kZ − Y kX ∀Z ∈ X ,
Y ∈X h

ku − Qh ukH1] (Ωc ) ≤ C2 inf ku − wkH1] (Ωc ) ∀u ∈ H1] (Ωc ).


w∈Mh

Let us introduce the following notations:

ρn1 := A(tn ) − Ph A(tn ), δ n1 := Ph A(tn ) − Anh , ¯


τ n1 := ∂A(t n ) − ∂t A(tn )

ρn2 := v(tn ) − Qh v(tn ), δ2n := Qh v(tn ) − vhn , ¯ n ) − ∂t v(tn ).


τ2n := ∂v(t
and

en1 := A(tn ) − Anh , en2 := v(tn ) − vhn .

Lemma 3. There exists a positive constant C, independent of h and ∆t:

n
X
kδ n1 k2X + kδ2n k2H1 (Ωc ) + ∆t kδ k1 k2X
]
k=1
( n h
)
X i
≤C kA0 − A0,h k2X + kρ01 k2X + kρn1 k2X + ∆t ¯ k k2
k∂ρ + kρk1 k2X + ¯ k k2 1
k∂ρ + kτ k1 k20,Ωc + kτ2k k20,Ωc .
1 X 2 H] (Ωc )
k=1

for n = 1, . . . , N .

Sketch of proof. The proof follows by using Lemma 1 and adapting the lines of [1, Lemma 3].

Theorem 1. If

A ∈ H1 (0, T ; X ) ∩ H2 (0, T ; L2 (Ω)), v ∈ H1 (0, T ; H1] (Ωc )) ∩ H2 (0, T ; L2 (Ωc )),

there exists a constant C > 0, independent of h and ∆t, such that

h i N
X
max ken1 k2X + ken2 k2H1 (Ωc ) + ∆t ken1 k2X
1≤n≤N ]
n=1
(  
≤C kA0 − A0,h k2X + max inf kA(tn ) − Zk2X + inf kv(tn ) − uk2H1 (Ωc )
1≤n≤N Z∈X h u∈Mh ]

N
X Z T  
+ ∆t inf kA(tn ) − Zk2X + inf k∂t A(t) − Zk2X dt
Z∈X h 0 Z∈X h
n=1
)
Z T   Z T
uk2H1 (Ωc ) 2 2 2
 
+ inf k∂t v(t) − dt + (∆t) k∂tt A(t)k0,Ωc + k∂tt u(t)k0,Ωc dt .
0 u∈Mh ]
0

Sketch of proof. The regularity assumptions on A and v allow us commute the time derivative with Ph and Qh , i.e,

∂t (Ph A(t)) = Ph (∂t A(t)) , ∂t (Qh v(t)) = Qh (∂t v(t)) .

Consequently, it is straightforward to deduce


N Z T   N Z T  
X
¯ n k2 ≤ C1 X
¯ n k2 1 C
k∂ρ1 X inf k∂t A(t) − Zk2X dt, k∂ρ 1 H (Ωc ) ≤ inf k∂t v(t) − uk2H1 (Ωc ) dt.
n=1
∆t 0 Z∈X h
n=1
] ∆t 0 u∈Mh ]

Therefore, the result follows by proceeding as in [1, Theorem 6.1].


Corollary 2. If there exists 0 < s ≤ 1 such that

A ∈ H1 (0, T ; X ∩ H1+s (Ω)3 ) ∩ H2 (0, T ; L2 (Ω)), v ∈ H1 (0, T ; H1] (Ωc ) ∩ H1+s (Ωc )) ∩ H2 (0, T ; L2 (Ωc )),

and A0,h = Πh (A0 ), where Πh : X ∩ H1+s (Ω)3 → X h is the standard finite element Lagrange interpolant, then there exists
a positive constant C, independent of h and ∆t, such that

h i N
X
ken1 k2X + ken2 k2H1 (Ωc ) + ∆t ken1 k2X ≤ C h2s + (∆t)2 .

max
1≤n≤N ]
n=1

Proof. Let Πh : H1] (Ωc ) ∩ H1+s (Ωc ) → Mh be the standard scalar finite element Lagrange interpolant. The result follows by
using previous theorem, the regularity assumptions on A and v, and the well-known approximation properties of Πh and Πh
(see, for instance, [8]).

Remark 2. Concerning this convergence result, it has to be noted that the spatial regularity of A (and in particular the regularity
of A0 ) is not ensured if Ω has reentrant corners or edges, namely, if it is a non-convex polyhedron. The result we have proved
here above ensures that the nodal finite element approximation is convergent either if the solution is regular (and this information
could be available even for a non-convex polyhedron Ω) or if Ω is a convex polyhedron. Actually, in [2, Section 5] is underlined
that if Ω is a convex polyhedron and B ∈ Hp (Ω)3 with 0 < p ≤ 1, the vector potential A (which is characterized by (11)–(12))
belongs to H1+s (Ω)3 for some 0 < s ≤ 1. Let us also note the assumption that Ω is convex is not a severe restriction, as in most
real-life applications ∂Ω arises from a somehow arbitrary truncation of the whole space and hence, reentrant corners and edges
of Ω can be easily avoided.

ACKNOWLEDGMENT
The authors would like to thank the University of Cauca for providing time for this work through research project VRI ID 3607.

REFERENCES
[1] R. Acevedo and G. Loaiza. A fully-discrete finite element approximation for the eddy currents problem. Ingenierı́a y Ciencia,
9(17): 111-145, 2013.
[2] R. Acevedo and R. Rodrı́guez, Analysis of the A, V − A − ψ potential formulation for the eddy current problem in a
bounded domain, Electronics Transactions on Numerical Analysis, 26, 270–284 (2007).
[3] C. Amrouche, C. Bernardi, C. Dauge and V. Girault, Vector potentials in three-dimensional non-smooth domains, Mathe-
matical Methods in the Applied Sciences, 21(9): 823–864 (1998)
[4] R. Araya, T. barrios, G. Gatica and N. Heuer, A posteriori error estimates for a mixed-FEM formulation of a non-linear
elliptic problem, Computer Methods in Applied Mechanics and Engineering, 191(21–22): 2317–2336 (2002).
[5] F. Bachinger, U. Langer and J. Schöberl. Numerical analysis of nonlinear multiharmonic eddy current problems. Nu-
merische Mathematik, 100(4): 593-616, 2005.
[6] O. Bı́ró and K. Preis, On the use of the magnetic vector potential in the finite element analysis of the three-dimensional
eddy currents. IEEE Transactions on Magnectics, 25(4): 3145-3159, 1989.
[7] O. Bı́ró and A. Valli, The Coulomb gauged vector potential formulation for the eddy-current problem in general geometry:
well-posedness and numerical approximation, Computer Methods in Applied Mechanics and Engineering, 196(13-16):
1890-1904, 2007.
[8] P. Ciarlet, The Finite Element Method for Elliptic Problems, SIAM, 2002.
[9] A. Ern and J. L. Guermond. Theory and Practice of Finite Elements. Springer, 2004.
[10] B. Heise, Analysis of a fully discrete finite element method for a nonlinear magnetic field problem. SIAM Journal on
Numerical Analysis, 31(3): 745-759, 1994.
[11] S. Reitzinger, B. Kaltenbacher and M. Kaltenbacher, A note on the approximation of B-H curves for nonlinear magnetic
field computations. In: Johannes Keppler Universität Linz, SFB “Numerical and Symbolic Scientific Computing”, SFB-
Report No 02-01, 2002.
[12] E. Zeidler, Nonlinear Functional Analysis and its Applications, II/B, Springer, New York, 1990.

You might also like