THE USE OF A REDUCED VECTOR POTENTIAL Ar FORMULATION

FOR THE CALCULATION OF IRON INDUCED FIELD ERRORS

Oszkár Bíró, Kurt Preis, Christian Paul

IGTE, Graz, Austria

Abstract
The application of the method of finite elements to computing the
magnetostatic field due to a given current density distribution in the
presence of ferromagnetic media is reviewed with the high precision needed
for the analysis of LHC magnets taken into account. Various formulations
in terms of either a magnetic scalar or vector potential are described. The
basic concepts of the method of finite elements are presented using both
node based and edge based elements. The formulation in terms of a reduced
magnetic vector potential is shown to be the best choice. It can be realized
with the aid of node based finite elements for two-dimensional models but
edge based elements are necessary when analyzing three-dimensional
arrangements.

1. INTRODUCTION
The magnetic field occurring in superconducting magnets including ferromagnetic iron parts can only
be computed numerically. Several potential formulations have been proposed in the past to serve as
the underlying boundary value problem [1-3]. These use either a reduced magnetic scalar potential or
a magnetic vector potential. Their merits and shortcomings will be discussed in the paper. The
conclusion arrived at is that, if highly permeable iron parts are present, the precision of formulations
based on a vector potential is higher.
The most versatile numerical technique for computing magnetic fields is the method of finite
elements (FEM). As pointed out below, node based or edge based finite elements can be employed
depending upon the potential formulation used. The scalar potential is best approximated with the aid
of nodal elements as is the single-component vector potential in two-dimensional problems. However,
if the vector potential is used in three-dimensional arrangements, the question of gauging arises. The
best method turns out to be to use an ungauged vector potential realized by edge elements [4].
Since the magnetic field in superconducting magnets such as the LHC dipoles has to be
computed with extremely high precision, it is desirable that the part of the field due to the
superconducting conductors be computed analytically using the Biot-Savart Law and only the part
due to the iron be obtained numerically with the aid of the method of finite elements. This means that
a reduced vector potential has to be used. This formulation, which ensures the high precision
required, will be presented in detail.

2. MAGNETOSTATIC FIELD
The differential equations of magnetostatic fields are the following Maxwell equations:

curl H = J (1)

div B = 0 (2)

where H is the magnetic field intensity, B is the magnetic flux density and J is the known current
density. The field quantities satisfy the following constitutive equations:

31
 B = µ (H) H or H = ν (B) B (3)

where µ is the permeability and ν the reluctivity, the reciprocal of the permeability. Due to the
saturation of iron, these material parameters depend on the magnetic field. For isotropic soft magnetic
media, wherein hysteresis is negligible, they can be assumed to be scalar quantities and monovalued
functions of the magnitude of the field. The closed domain in which the magnetic field is to be
calculated will be denoted by Ω.
The field quantities B and H satisfy boundary conditions on the boundary of Ω. Two types of
boundary conditions cover all practical cases. They are prescribed on two disjunct parts ΓB and ΓH of
the boundary with the union of ΓB and ΓH forming the entire boundary.
On the part ΓB of the boundary, the normal component of the magnetic flux density is known.
In many cases this value is zero, as on symmetry planes parallel to the field. Since, in order to employ
the method of finite elements, a closed domain Ω has to be assumed, artificial far boundaries are
frequently introduced. These far boundaries may also be part of ΓB with the normal component of B
vanishing. In some special problems, the distribution of Bnormal can be estimated along a physical
surface. As an example, it can often be assumed that no flux leaves the outer boundary of an iron
structure completely surrounded by air or that the flux distribution in the air gap of an electrical
machine is sinusoidal. All these boundary conditions can be written in the form

B ⋅ n = −b on ΓB (4)

where n is the outer unit normal vector on ΓB and b can interpreted as a fictitious magnetic surface
charge density. (The negative sign in (4) implies that positive values of b correspond to positive
surface charges.)

On the part ΓH of the boundary, the tangential component of the magnetic field intensity is
known. In many cases this value is zero, as on symmetry planes perpendicular to the field. Far
boundaries may also be part of ΓH with the tangential component of H vanishing. In some special
problems, the distribution of Htangential can be estimated along a physical surface. For example, it can
often be assumed that the field enters highly (infinitely) permeable iron structures at right angle or
that the tangential component of H is determined by a surface current flowing on the surface of an
infinitely permeable magnetic pole. All these boundary conditions can be written in the form

H × n = K on ΓH (5)

where n is the outer unit normal vector on ΓH and K can interpreted as a real or fictitious electric
surface current density.

The interface conditions on any surface between two regions with different magnetic properties
are the continuity of Bnormal and of Htangential. Denoting the outer unit normal vectors of the two abutting
regions Ω1 and Ω2 by n1 and n2 and using the indices 1 and 2 to denote the field quantities in the two
regions, the interface conditions on the interface Γ12 can be written as

B1 ⋅ n1 + B 2 ⋅ n 2 = 0  (6a)
 on Γ12 .
H 1 × n 1 + H 2 × n 2 = 0 (6b)

In case ΓH is a connected surface, no further conditions are necessary to define the static
magnetic field. If, however, ΓH consists of nH+1 disjunct parts ΓH0, ΓH1, ..., ΓHnH, then either the nH
magnetic voltages of between ΓH1, ..., ΓHnH and ΓH0 must be defined as

32
 ∫ H ⋅ dl = U
C Hi
mi , i = 1, 2, ..., nH (7)

where CHi is an arbitrary curve connecting ΓHi and ΓH0, or the nH magnetic fluxes of the surfaces ΓH1, ...,
ΓHnH have to be given as

∫ B ⋅ ndΓ = Ψ ,
ΓHi
i i = 1, 2, ..., nH. (8)

3. POTENTIAL FORMULATIONS
The solution of the differential equations (1) and (2) with the constitutive equation (3), the boundary
conditions (4) and (5), the interface conditions (6) as well as the integral conditions (7) or (8) is
unique [5]. Equivalent boundary value problems involving second order elliptic differential equations
as well as Dirichlet and Neumann boundary conditions can be set up in terms of potential functions.
These boundary value problems lend themselves well to numerical solution with the aid of the
method of finite elements. The various potential formulations will be reviewed in the following.

3.1 Reduced scalar potential

Since the curl of the magnetic field intensity is, in general, nonzero, it cannot always be written as the
gradient of a scalar potential function. If, however, a function T is found satisfying

curlT = J in Ω, (9)

then H-T is curl free and the magnetic field intensity can be written as

H = T − gradΦ (10)

where Φ is the reduced magnetic scalar potential.

The choice (10) automatically satisfies Ampere’s Law (1), so Maxwell’s equation (2) remains
to be solved. Taking account of the material relationship (3), it has the form

− div( µgradΦ ) = −div( µT) in Ω. (11)

This is a generalized Laplace-Poisson equation.

The boundary condition (4) becomes

∂Φ
µn ⋅ gradΦ = µ = µT ⋅ n + b on ΓB, (12)
∂n

a nonhomogeneous Neumann boundary condition. Expressing the condition (5) with the aid of the
scalar potential,

gradΦ × n = T × n − K on ΓH (13)

is obtained. Let us assume that the current density J has no normal component on ΓH and that the
integral quantities specified if ΓH is composed of several nonconnected parts are the magnetic
33
voltages as in Eq. (7). Then, choosing the value of Φ to be the magnetic voltage Umi defined in (7) at
an arbitrary point P0i in ΓHi, i = 1, 2, ..., nH, Um0 = 0, the value of Φ at any point P in ΓHi is obtained as
the sum of Umi and of the integral of the tangential component of gradΦ along some curve CPi
connecting P0i to P. The tangential component of gradΦ can be obtained from Eq. (13) as

∂Φ
= n × (gradΦ × n) = n × (T × n) + K × n on ΓH. (14)
∂t

Hence, the boundary condition on ΓH is

Φ ( P) = U mi + ∫ [n × (T × n) + K × n] ⋅ dl = Φ 0 ( P ) , P ∈ ΓHi , (15)
C Pi

a Dirichlet boundary condition. If fluxes of the form (8) are the specified as integral quantities, then
the reduced scalar potential formulation cannot be employed directly.
The boundary value problem consisting of the differential equation (11), the Neumann
boundary condition (12) and the Dirichlet boundary condition (15) can be cast in a weak form
facilitating the application of the method of finite element to its numerical solution. Let us consider
all functions satisfying the Dirichlet boundary condition (15). From among these, the solution of the
boundary value problem is the function Φ fulfilling

∫ w[−div( µgradΦ)]dΩ + ∫ w(µn ⋅ gradΦ)dΓ = ∫ w[−div(µT)]dΩ + ∫ w( µT ⋅ n + b)dΓ

Ω ΓB Ω ΓB
(16)

with w being an arbitrary weighting function that obeys the homogeneous counterpart of the Dirichlet
boundary condition (15):

w = 0 on ΓH. (17)

Using the identities

∫ w[−div( µgradΦ)]dΩ = ∫ gradw ⋅ µgradΦdΩ − ∫ w(µn ⋅ gradΦ)dΓ ,

Ω Ω ΓB + ΓH
(18)

Ω
∫ w[−div(µT)]dΩ = ∫ gradw ⋅ µTdΩ − ∫ w(µT ⋅ n)dΓ
Ω ΓB + ΓH
(19)

as well as the boundary condition (17), the following can be stated:

The solution of the boundary value problem (11), (12), (15) is the function Φ satisfying the
Dirichlet boundary condition (15) if the weak form

∫ gradw ⋅ µgradΦdΩ = ∫ gradw ⋅ µTdΩ + ∫ wbdΓ

Ω Ω ΓB
(20)

holds for any function w satisfying Eq. (17).

34
 In order to satisfy the interface condition (6a), it is sufficient that the scalar potential as well as
the tangential component of T be continuous. The interface condition (6b) is included in the weak
form provided w is continuous along Γ12. This can be seen by adding the two integrals over Γ12
corresponding to (6b) to the left hand side of (16). Applying the identities (18) and (19) over Ω1 and
Ω2, the form (20) is obtained.
Several options are open for the choice of the function T satisfying Eq. (9). The most
straightforward one is the Biot-Savart field HS computed from the current density as

1 J (Q) × e QP
H S ( P) =
4π ∫
ΩQ
2
rQP
dΩ Q (21)

where eQP is the unit vector pointing from the source point Q to the field point P, rQP is the distance
between Q and P and ΩQ is the domain where J is nonzero. The curl of HS is obviously J, so it
satisfies Eq. (9) and is thus a valid choice as T. In highly permeable media, the magnitude of H is
frequently much less than the magnitude of HS. This means that using HS as T can result in large
cancellation errors when computing H from Eq. (10). These cancellation errors are ruinous if Φ is
computed numerically, e.g. approximated by piecewise continuous functions with discontinuous
derivatives as in the method of finite elements, and HS is simultaneously computed analytically as a
smooth function. One method to avoid these errors is using a total scalar potential in ferromagnetic
regions assumed to be current free [1]. An alternative is to interpolate HS with the aid similar
functions as those used for the approximation of gradΦ [3]. The choice of using HS as T is, naturally,
not the only possibility. Some useful options in the context of the method of finite elements can be
found in [6] and [3].
The scalar potential formulation of magnetostatic fields is appealing since it offers the most
economic description in terms of unknown functions. In view of the fact, however, that the quantity
derived directly from the potential function is the magnetic field intensity, the error of the flux
density is much higher in ferromagnetic iron regions than in nonferromagnetic air domains especially
if the field runs mainly parallel to the iron/air interface and hence H has about the same value in both
regions [7]. This fact suggest that formulations yielding the magnetic flux density directly from the
potentials perform better in problems involving highly permeable parts.

3.2 Vector potential

Since the divergence of the magnetic flux density is zero, it can be written as the curl of a magnetic
vector potential function A:
B = curlA in Ω. (22)

The choice (22) automatically satisfies Maxwell’s equation (2), so Ampere’s Law (1) remains
to be solved. Taking account of the material relationship (3), it has the form

curl (νcurlA) = J in Ω. (23)

This is a second order partial differential equation.

The boundary condition (5) becomes

νcurlA × n = K on ΓH, (24)

a nonhomogeneous Neumann boundary condition. Expressing the condition (4) with the aid of the
vector potential,

35
 n ⋅ curlA = b on ΓB (25)

is obtained. The normal component of the curl of A is completely determined by the tangential
component of the vector potential. Therefore it is possible to choose a function a so that the Dirichlet
boundary condition

n × A = a on ΓB (26)

implies (25). Naturally, several possible functions a exist. All they have to fulfill are the two
conditions

diva = div(n × A) = n ⋅ curlA = b on ΓB (27)

and

∫ (a × n) ⋅ dl = ∫ A
C Hi C Hi
tangential ⋅ dl = ∫ A ⋅ dl = ∫ n ⋅ curlAdΓ = Ψ
C Hi ΓHi
i , i = 1, 2, ..., nH (28)

where CHi is the curve bounding the surface ΓHi and as such separates ΓHi from ΓΒ. The satisfaction of
Eq. (28) ensures the fulfillment of the integral conditions (8). If magnetic voltages of the form (7) are
the specified integral quantities, then the vector potential formulation cannot be employed directly.
The solution of the boundary value problem consisting of the differential equation (23), the
Neumann boundary condition (24) and the Dirichlet boundary condition (26) is not unique. The
gradient of any scalar function can be added to any of its solutions if the scalar function is constant on
ΓΒ. This lack of uniqueness can be eliminated by modifying the boundary value problem to include
the Coulomb gauge on the vector potential [2]. In this case, the differential equation (23) is replaced
by

curl (νcurlA) − grad (νdivA) = J in Ω, (23a)

and the boundary conditions are supplemented by the following two conditions:

A ⋅ n = 0 on ΓH, (24a)

νdivA = 0 on ΓB. (26a)

The boundary value problem (23a), (24), (24a), (26), (26a) has a unique solution satisfying the
Coulomb gauge

νdivA = 0 in Ω (29)

and hence also the differential equation (23).

To obtain the weak form of the ungauged, nonunique boundary value problem (23), (24), (26),
the satisfaction of the Dirichlet boundary condition (26) is assumed and (23), (24) are written by
requiring that

36
 ∫ w ⋅ [curl (νcurlA)]dΩ + ∫ w ⋅ (νcurlA × n)dΓ = ∫ w ⋅ JdΩ + ∫ w ⋅ KdΓ
Ω ΓH Ω ΓH
(30)

is satisfied with w being an arbitrary vector weighting function that obeys the homogeneous
counterpart of the Dirichlet boundary condition (26):

n × w = 0 on ΓB. (31)

Using the identity

∫ w ⋅ [curl (νcurlA)]dΩ = ∫ curlw ⋅νcurlAdΩ − ∫ w ⋅ (νcurlA × n)dΓ

Ω Ω ΓB + ΓH
(32)

as well as the boundary condition (31), the following can be stated:

A solution of the boundary value problem (23), (24), (26) is any function A satisfying the
Dirichlet boundary condition (26) if the weak form

∫ curlw ⋅νcurlAdΩ = ∫ w ⋅ JdΩ + ∫ w ⋅ KdΓ

Ω Ω ΓH
(33)

holds for any function w satisfying Eq. (31). If the current density J and the surface current density K
are described with the aid of a function T satisfying (9) and

T × n = K on ΓH, (34)

then, using the identity

∫ w ⋅ curlTdΩ = ∫ curlw ⋅ TdΩ − ∫ w ⋅ (T × n)dΓ

Ω Ω ΓB + ΓH
(35)

and the boundary condition (31), the weak form (33) can be rewritten as

∫ curlw ⋅νcurlAdΩ = ∫ curlw ⋅ TdΩ .

Ω Ω
(36)

In order to satisfy the interface condition (6b), it is sufficient that the tangential component of
A be continuous. The interface condition (6a) is included in the weak form provided the tangential
component of w is continuous along Γ12. This can be seen by adding the two integrals over Γ12
corresponding to (6a) to the left hand side of (30). Applying the identity (32) over Ω1 and Ω2, the
form (33) is obtained.
To obtain the weak form of the gauged, unique boundary value problem (23a), (24), (24a), (26),
(26a), one has to assume the satisfaction of the Dirichlet boundary conditions (24a) and (26) whereas
(23a), (24) and (26a) are written by requiring that

37
 ∫ w ⋅ [curl (νcurlA) − grad (νdivA)]dΩ + ∫ w ⋅ (νcurlA × n)dΓ + ∫ w ⋅ nνdivAdΓ = ∫ w ⋅ JdΩ + ∫ w ⋅ KdΓ
Ω ΓH ΓB Ω ΓH

(37)

is satisfied with w being an arbitrary vector weighting function that obeys the homogeneous
counterparts of the Dirichlet boundary conditions (24a) and (26):

w ⋅ n = 0 on ΓH, (38)

n × w = 0 on ΓB. (39)

Using the identities (32) and

∫ w ⋅ [− grad (νdivA)]dΩ = ∫ divwνdivAdΩ − ∫ w ⋅ nνdivAdΓ

Ω Ω ΓB + ΓH
(40)

as well as the boundary conditions (38) and (39), the following can be stated:
The solution of the boundary value problem (23a), (24), (24a), (26), (26a) is the function A
satisfying the Dirichlet boundary conditions (24a), (26) if the weak form

∫ (curlw ⋅νcurlA + divwνdivA)dΩ = ∫ w ⋅ JdΩ + ∫ w ⋅ KdΓ

Ω Ω ΓH
(41)

holds for any function w satisfying Eqs. (38) and (39).

It is easy to verify that, besides the interface condition (6a), this weak also form implies the
condition

n1ν 1divA1 + n 2ν 2 divA 2 = 0 on Γ12. (42)

It will be shown in section 4 that the application of the method of finite elements ensures the
approximate satisfaction of the weak forms. This means that a certain error in the fulfillment of the
Coulomb gauge (29) is invariably present. Along an iron/air interface, the condition (42) implies that
this error is much higher in the ferromagnetic region (where ν is low) than in the air domain (where ν
may be several thousand times higher). A large error in the satisfaction of the Coulomb gauge results
in even larger errors in fulfilling Ampere’s Law in iron, since the additional term grad(νdivA) in
(23a) is far from zero [8]. All in all, this feature of the gauged formulation makes it unsuitable for
solving problems involving ferromagnetic materials. Therefore, in the following, the ungauged
boundary value problem (23), (24), (26) will be considered only.

4. THE METHOD OF FINITE ELEMENTS

The method of finite elements requires the problem region Ω to be discretized into elementary
domains called finite elements forming a mesh.. The elements have simple geometrical forms such as
triangles or quadrilaterals in two-dimensional problems or tetrahedra, hexahedra, prisms, etc. in three-
dimensional arrangements. The elements are defined by nodes and scalar interpolating functions
called shape functions are associated with each node in each element. The shape functions are low
order polynomials. Scalar potentials can be interpolated with the aid of the shape functions providing
approximations which are continuous on the interfaces between finite elements. Such nodal finite
38
elements will be discussed in section 4.1. Besides nodes, edges can also be defined within finite
elements and vector shape functions are then associated with each edge in each element. These edge
shape functions are low order vector polynomials. They can be used to interpolate vector functions.
The approximations thus generated are vector functions whose tangential components are continuous
across element interfaces but whose normal components are, in general, discontinuous. They are,
therefore, suitable for approximating vector potentials. Edge finite elements will be treated in section
4.2. A detailed exposition of the method of finite elements can be found e.g. in [9].

4.1 Node based elements

(e )
An element node shape function N k j (k = 1, 2, ..., nn(e ) ) is associated with each of the nn(e ) nodes of
the j-th finite element. They are usually defined in a local coordinate system and are low order
polynomials of the local coordinates ξ, η, ζ. They are constructed to obey the equations

1 at the local node k ,

N k j (ξ ,η , ζ ) = 
(e )
k = 1, 2, ..., nn(e ) . (43)
î 0 at all other local nodes,

The element node shape functions may also serve to establish a transformation between local
and global coordinates. With x, y and z denoting the global coordinates, this transformation has the
following form within the j-th element:

nn( e ) nn( e ) nn( e )

x(ξ ,η , ζ ) = ∑x N
k =1
k
(e j )
k (ξ ,η , ζ ) , y (ξ ,η , ζ ) = ∑y N
k =1
k
(e j )
k (ξ ,η , ζ ) , z (ξ ,η , ζ ) = ∑z N
k =1
k
(e j )
k (ξ ,η , ζ ) .

(44)

where xk, yk and zk are the global coordinates of the k-th local node.
Since the nodes of neighboring elements coincide, the number of the global nodes is less than
the product of the number of elements and nn(e ) . It will be denoted by nn. A global node shape
function Ni (i = 1, 2, ..., nn) is associated with each of the global nodes and is defined as

 N j ( x, y, z ) in the j - th element, if the global node i coincides with its local node k ,
(e )

N i ( x, y , z ) =  k
î0 in an element, if the global node i does not coincide with any of its local nodes .

(45)

The global node shape functions are continuous in Ω since they have the same variation in both
elements along any interface between two elements. They have a property similar to Eq. (43) with
respect to the global nodes:

1 at the global node i,

N i ( x, y, z ) =  i = 1, 2, ..., nn. (46)
î 0 at all other global nodes,

The nn global node shape functions are linearly independent, but there is a linear
interdependence among their gradients. In fact, the sum of all nodal basis functions is 1:

39
 nn

∑Ni =1
i = 1, (47)

as it is obvious from the fact that the function identically equal to 1 can be exactly represented with
the aid of node shape functions (the constant is the lowest order polynomial) and all its nodal values
are 1. The following is obtained after taking the gradient of (47):

nn

∑ gradN
i =1
i = 0. (48)

This means that the maximal number of linearly independent gradients of the global node shape
functions is nn-1, i.e. the number of tree edges in the graph defined by the finite element mesh. It
follows from the linear independence of the global node shape functions that nn-1 of their gradients
are in fact linearly independent.
Let us now consider the numerical solution of the weak form (20) of the reduced scalar
potential formulation of magnetostatic fields. Let the global nodes in the finite element mesh which
are not on the surface ΓH (where a Dirichlet boundary condition is prescribed) obtain the order
numbers 1, 2, ..., n and those on ΓH the order numbers n+1, n+2, ..., nn. Let further Φi denote the value
of the reduced magnetic scalar potential in the global node i. Obviously, the values Φi, i = n+1, n+2,
..., nn are known from the Dirichlet boundary condition (15) and, due to the interpolatory property
(46) of the global node shape functions, the expression

nn

Φ D ( x, y , z ) = ∑ Φ N ( x, y , z )
i = n +1
i i (49)

is a known function approximately satisfying Eq. (15). Furthermore, the functions Ni, i = 1, 2, ..., n
satisfy the homogeneous Dirichlet boundary condition (17) required of the weighting functions.
Therefore, an approximation of the scalar potential in the form

nn

∑ Φ N ( x, y , z ) = Φ ∑ Φ N ( x, y , z )
n

Φ ( x, y , z ) ≈ Φ ( x, y , z ) =
(n)
i i D ( x, y , z ) + i i (50)
i =1 i =1

is suitable for a numerical solution based on the weak form (20). Indeed, it satisfies the Dirichlet
boundary condition (15) independent of the choice of the n unknown nodal potential values Φi, i = 1,
2, ..., n.
The relevant numerical method is called Galerkin’s procedure and is constituted by writing the
weak form (20) with the scalar potential function replaced by the approximation (50) and using the
functions Ni, i = 1, 2, ..., n as weighting functions. This leads to a system of algebraic equations for
the n unknowns:

∫ gradN ⋅ µgradΦ dΩ =
∫ gradN ⋅ µTdΩ + ∫ N bdΓ , i = 1, 2, ..., n,
( n)
i i i (51)
Ω Ω ΓB

or, with the known quantities brought to the right hand side:

40
 ∑ Φ ∫ gradN ⋅ µgradN dΩ
n

j i j
j =1 Ω

=
∫ gradN ⋅ µTdΩ + ∫ N bdΓ − ∫ gradN ⋅ µgradΦ
Ω
i
ΓB
i
Ω
i D dΩ , i = 1, 2, ..., n. (52)

The matrix of this equations system is obviously symmetric and also sparse, since the support of the
global node shape functions extends over a few finite elements only. The matrix is also positive
definite. The system can be solved advantageously with the aid of iterative techniques. The most
widely spread method is that of preconditioned conjugate gradients [10].

4.2 Edge based elements

(e )
An element edge shape function N k j (k = 1, 2, ..., ne(e ) ) is associated with each of the ne(e ) edges of
the j-th finite element. They are usually defined in a local coordinate system and are low order vector
polynomials of the local coordinates ξ, η, ζ and of their gradients. This results in the important
property that the gradients of the node shape functions are in the space spanned by the edge shape
functions or, in other words, the gradients of the node shape functions can be written as linear
combinations of the edge shape functions. The latter are constructed to obey the equations

1 if l = k ,
∫N ⋅ dl = 
(e j )
k k = 1, 2, ..., ne(e ) . (53)
edgel
î 0 otherwise,

where edgel is the l-th edge of the element.

The transformation between local and global coordinates is established with the aid of the
element node shape functions as in Eq. (44).
Since the edges of neighboring elements coincide, the number of the global edges is less than
the product of the number of elements and ne(e ) . It will be denoted by ne. A global edge shape function
Ni (i = 1, 2, ..., ne) is associated with each of the global edges and is defined as

N ( e j ) ( x, y, z ) in the j - th element, if the global edge i coincides with its local edge k ,
N i ( x, y , z ) =  k
î0 in an element, if the global edge i does not coincide with any of its local edges .

(54)

The tangential components of the global edge shape functions are continuous in Ω since they have the
same variation in both elements along any interface between two elements. They have a property
similar to Eq. (53) with respect to the global edges:

1 if j = i,
∫ N ⋅ dl = î 0 otherwise,
edge j
i i = 1, 2, ..., ne. (55)

The ne global edge shape functions are linearly independent, but there are linear interdependencies
among their curls. Indeed, since the gradients of the nodal basis functions are in the function space
spanned by the edge basis functions, we have the following nn-1 linearly independent relations:
41
 ne
gradN i = ∑ cik N k , i=1,2,...,nn-1 (56)
k =1

where

ne

∑c
k =1
2
ik > 0 , i=1,2,...,nn-1. (57)

Taking the curl of each of the equations in (56) results in

ne

∑c
k =1
ik curlN k = 0 , i=1,2,...,nn-1. (58)

Together with (57) and with the linear independence of the equations in (56), this implies that the
maximal number of linearly independent curls of the edge basis functions is ne -(nn-1), i.e. the number
of cotree edges in the graph of the finite element mesh. Since there are no more linearly independent
gradients in the space spanned by the edge basis functions than nn-1, not less than ne -(nn-1) of the
curls of the edge basis functions are linearly independent.
Let us now consider the numerical solution of the weak form (33) of the ungauged vector
potential formulation of magnetostatic fields. Let the global edges in the finite element mesh which
are not on the surface ΓB (where a Dirichlet boundary condition is prescribed) obtain the order
numbers 1, 2, ..., n and those on ΓB the order numbers n+1, n+2, ..., ne. Let further Αi denote the
integral of the magnetic vector potential over the global edge i. Obviously, the values Αi, i = n+1,
n+2, ..., ne are known from the Dirichlet boundary condition (26) and, due to the interpolatory
property (53) of the global edge shape functions, the expression

ne
A D ( x, y , z ) = ∑ A N ( x, y , z )
i = n +1
i i (59)

is a known function approximately satisfying Eq. (26). Furthermore, the functions Ni, i = 1, 2, ..., n
satisfy the homogeneous Dirichlet boundary condition (31) required of the weighting functions.
Therefore, an approximation of the vector potential in the form

ne

∑ ∑ A N ( x, y , z )
n

A ( x, y , z ) ≈ A ( x, y , z ) =
(n)
Ai N i ( x, y, z ) = A D ( x, y, z ) + i i (60)
i =1 i =1

is suitable for a numerical solution based on the weak form (33). Indeed, it satisfies the Dirichlet
boundary condition (26) independent of the choice of the n unknown integral values Αi, i = 1, 2, ..., n.
The application of Galerkin’s method is constituted by writing the weak form (33) with the
vector potential function replaced by the approximation (60) and using the functions Ni, i = 1, 2, ..., n
as weighting functions. This leads to a system of algebraic equations for the n unknowns:

∫ curlN ⋅νcurlA ∫ ∫
dΩ = N i ⋅ JdΩ + N i ⋅ KdΓ , i = 1, 2, ..., n,
(n)
i (61)
Ω Ω ΓH
42
or, with the known quantities brought to the right hand side:

∑ A ∫ curlN ⋅νcurlN dΩ = ∫ N ⋅ JdΩ + ∫ N ⋅ KdΓ − ∫ curlN ⋅νcurlA dΩ ,

n

j i j i i i D
j =1 Ω Ω ΓH Ω

i = 1, 2, ..., n. (62)

The matrix of this equations system is obviously symmetric and also sparse, since, similarly to the
node shape functions, the support of the global edge shape functions extends over a few finite
elements only. The matrix is also positive semidefinite, i.e. all its eigenvalues are nonnegative, but
some of them are zero. The singularity of the matrix follows immediately from the linear
interdependencies between the curls of the global edge shape functions written in Eq. (58). Since the
method of conjugate gradients can cope efficiently with positive semidefinite (singular) matrices [11]
provided the right hand side of the equations system is consistent, the robustness of the numerical
scheme is ensured. Note that this would not be the case if node shape functions were used to
approximate each component of the vector potential: the matrix in the equations system (62) were
then not singular but, because of the small eigenvalues approximating zero, extremely ill conditioned.
The form (62) of the Galerkin equations does not ensure the consistence of the right hand side.
Using the weak form (36) instead of (33), however, results in the equations system

∑ A ∫ curlN ⋅νcurlN dΩ = ∫ curlN ⋅ TdΩ − ∫ curlN ⋅νcurlA

n

j i j i i D dΩ , i = 1, 2, ..., n. (63)
j =1 Ω Ω Ω

In this form, the right hand side is obviously consistent, since the same linear interdependence among
its elements is present as among the rows of the left hand side. Consequently, the form (63) of the
Galerkin equations must be used in the numerical realization.

5. REDUCED VECTOR POTENTIAL

The total vector potential formulation described in section 3.2 with its numerical solution by means of
edge finite elements presented in section 4.2 has the disadvantage that the shape of the coils has to be
exactly modeled by finite elements. If this is not the case, then the precision of the right hand side of
the Galerkin equations becomes very low as it can be seen in Eq. (62). Indeed, the numerical
integration of the product of the edge shape functions and of the current density cannot be carried out
precisely if J is discontinuous within the finite elements. In addition, it is desirable that the part of the
field due to the conductors be computed analytically using the Biot-Savart Law and only the part due
to the iron be obtained numerically with the aid of the method of finite elements.
The necessity of representing the shape of coils by the finite element mesh can be avoided by
introducing a reduced vector potential Ar [3] as

B = µ 0 H S + curlA r in Ω (64)

where HS is the Biot-Savart field defined in (21). Also, µ0HS is the magnetic field due to the coils in
free space and, hence, curlAr is the field resulting from the presence of iron.
The choice (64) automatically satisfies Maxwell’s equation (2) since the divergence of the
Biot-Savart field is zero. Ampere’s Law (1) remains to be solved. Taking account of the material
relationship (3) and of the fact that the curl of HS is J, it has the form

43
 curl (νcurlA r ) = curlH S − curl (νµ 0 H S ) in Ω. (65)

This second order partial differential equation is similar to Eq. (23). In air regions, the right hand side
is obviously zero.
The boundary condition (5) becomes

νcurlA r × n = K − νH S × n on ΓH, (66)

a nonhomogeneous Neumann boundary condition. Expressing the condition (4) with the aid of the
reduced vector potential,

n ⋅ curlA r = b − µ 0 n ⋅ H S on ΓB (67)

is obtained. This is equivalent to the Dirichlet boundary condition

n × A r = a on ΓB (68)

where a satisfies two conditions similar to (27) and (28).

To obtain the weak form of the boundary value problem (65), (66), (68), the satisfaction of the
Dirichlet boundary condition (68) is assumed and (65), (66) are written by requiring that

∫ w ⋅ [curl (νcurlA
Ω
r )]dΩ + ∫ w ⋅ (νcurlA r × n)dΓ
ΓH

= ∫ w ⋅ [curlH S − curl (νµ 0 H S )]dΩ + ∫ w ⋅ (K − νµ 0 H S × n)dΓ (69)

Ω ΓH

is satisfied with w being an arbitrary vector weighting function that obeys the homogeneous Dirichlet
boundary condition (31)
Using the identities

∫ w ⋅ [curl (νcurlA
Ω
r )]dΩ = ∫ curlw ⋅νcurlA r dΩ −
Ω
∫ w ⋅ (νcurlA
ΓB + ΓH
r × n ) dΓ , (70)

∫ w ⋅ [curl (νµ H
Ω
0 S )]dΩ = ∫ curlw ⋅νµ 0 H S dΩ −
Ω
∫ w ⋅ (νµ H
ΓB + ΓH
0 S × n ) dΓ , (71)

∫ w ⋅ curlH
Ω
S dΩ = ∫ curlw ⋅ H S dΩ −
Ω
∫ w ⋅ (H
ΓB + ΓH
S × n ) dΓ , (72)

assuming similarly to (34) that

H S × n = K on ΓH (73)

and using the boundary condition (31), the following can be stated:
44
 A solution of the boundary value problem (65), (66), (68) is any function Ar satisfying the
Dirichlet boundary condition (68) if the weak form

Ω
∫ curlw ⋅νcurlA dΩ = ∫ curlw ⋅ (H
r
Ω
S − νµ 0 H S )dΩ (74)

holds for any function w satisfying Eq. (31).

In order to solve the weak form (74) by Galerkin’s method using edge finite elements, let the
global edges in the finite element mesh which are not on the surface ΓB (where a Dirichlet boundary
condition is prescribed) obtain the order numbers 1, 2, ..., n and those on ΓB the order numbers n+1,
n+2, ..., ne. Let further Αi denote the integral of the reduced magnetic vector potential over the global
edge i. Obviously, the values Αi, i = n+1, n+2, ..., ne are known from the Dirichlet boundary condition
(68) and, due to the interpolatory property (53) of the global edge shape functions, the expression

ne
A D ( x, y , z ) = ∑ A N ( x, y , z )
i = n +1
i i (75)

is a known function approximately satisfying Eq. (68). Furthermore, the functions Ni, i = 1, 2, ..., n
satisfy the homogeneous Dirichlet boundary condition (31) required of the weighting functions.
Therefore, an approximation of the reduced vector potential in the form

ne n
A r ( x, y, z ) ≈ A r( n ) ( x, y, z ) = ∑ Ai N i ( x, y , z ) = A D ( x, y , z ) + ∑ Ai N i ( x, y, z ) (76)
i =1 i =1

is suitable for a numerical solution based on the weak form (74). Indeed, it satisfies the Dirichlet
boundary condition (68) independent of the choice of the n unknown integral values Αi, i = 1, 2, ..., n.
The application of Galerkin’s method is again constituted by writing the weak form (74) with
the vector potential function replaced by the approximation (76) and using the functions Ni, i = 1, 2,
..., n as weighting functions. This leads to a system of algebraic equations for the n unknowns:

∫ curlN
Ω
i ⋅νcurlA (rn ) dΩ = ∫ curlN i ⋅ (H S − νµ 0 H S )dΩ , i = 1, 2, ..., n,
Ω
(77)

or, with the known quantities brought to the right hand side:

∑ A ∫ curlN ⋅νcurlN dΩ = ∫ curlN ⋅ (H

j =1
j
Ω
i j
Ω
i S − νµ 0 H S )dΩ − ∫ curlN i ⋅νcurlA D dΩ ,
Ω

i = 1, 2, ..., n. (78)

The matrix of this equations system is the same as in Eq. (62) obtained in the case of the total vector
potential. The consistence of the right hand side is obvious, so the singularity of the matrix does not
impair the robustness of the numerical scheme.

45
6. CONCLUSIONS
The numerical solution of magnetostatic fields by means of the finite element method can be based
either on a reduced magnetic scalar potential or on a magnetic vector potential. The first option can be
realized by node based finite elements and the second one by edge based ones. The precision of the
vector potential formulation is superior to that of the method based on a scalar potential in highly
permeable iron parts. The use of edge based finite elements for the numerical solution of the
boundary value problem in terms of an ungauged vector potential leads to an equations system with a
singular matrix. Choosing a suitable form of the right hand side, it can be made to be consistent and,
hence, the iterative solution of the equations system is robust. Taking account of the field of the coils
in free space with the aid of the Biot-Savart Law, a reduced magnetic vector potential can be
introduced. This eliminates the necessity of modeling the shape of the coils by the finite element
mesh and results in high precision since the finite element solution represents the iron induced fields
only. Consequently, the reduced vector potential formulation is especially suitable for the analysis of
superconducting magnets including iron parts such as the LHC magnets.

REFERENCES
[1] J. Simkin and C.W. Trowbridge, On the use of the total scalar potential in the numerical
solution of field problems in electromagnetics, Int. J. Numer. Meth. Eng., 14 (1979) 423.
[2] K. Preis, O. Biro, C. A. Magele, W. Renhart, K.R. Richter and G. Vrisk, Numerical analysis
of 3D magnetostatic fields, IEEE Trans. Magn., 27 (1991) 3798.
[3] O. Biro, K. Preis, G. Vrisk, K.R. Richter and I. Ticar, Computation of 3D magnetostatic
fields using a reduced scalar potential, IEEE Trans. Magn., 29 (1993) 1329.
[4] O. Biro, K. Preis and K.R. Richter, On the use of the magnetic vector potential in the nodal
and edge finite element analysis of 3D magnetostatic fields, IEEE Trans. Magn., 32 (1996)
651.
[5] O. Biro and K.R. Richter, CAD in Electromagnetism, in P. W. Hawkes (ed.), Advances in
Electronics and Electron Physics, (Academic Press) 82 (1991) 96.
[6] J.P. Webb and B. Forghani, A single scalar potential method for 3D edge elements using
edge elements, IEEE Trans. Magn., 25 (1989) 4126.
[7] O. Biro, Ch. Magele and G. Vrisk, Solution of TEAM benchmark problem #13 (3D nonlinear
magnetostatic model), ACES Journal, 8 (1993) 216.
[8] O. Biro, Solution of TEAM benchmark problem #10 (Steel plates around a coil), ACES
Journal, 8 (1993) 203.
[9] O.C. Zienkiewicz and R.L. Taylor, The finite element method, Fourth Edition, McGraw-Hill
Book Company
[10] D.S. Kershaw, The incomplete Cholesky-conjugate gradient method for iterative solution of
systems of linear equations, J. Comput. Phys, 26 (1978) 43.
[11] R.W. Freund and M. Hochbruck, On the use of two QMR algorithms for solving singular
systems and applications in Markov chain modeling, Numerical Linear Algebra with
Applications, 1 (1994) 403.

46