You are on page 1of 16

Numerical investigation of optimal control of induction

heating processes
Olivier Bodart
a
, Anne-Valerie Boureau
b
, Rachid Touzani
a,
*
a
Laboratoire de Math ematiques Appliqu ees, Universit e Blaise Pascal, (Clermont-Ferrand) and CNRS (UMR 6620),
63177 Aubi ere Cedex, France
b
Logatique, 50 rue Marcel-Dassault, 92100 Boulogne-Billancourt, France
Received 29 October 1998; received in revised form 16 October 2000; accepted 4 December 2000
Abstract
We consider optimal control problems arising in induction heating processes. We are mainly concerned with two
classes of these processes: uniform heating and metal hardening. The cost functions are chosen according to these
classes. The control parameters are the inductor shape (assumed to be thin), the frequency, the current voltage and the
heating duration. The induction heating model is a two-dimensional cartesian geometry. The numerical scheme is given
as well as numerical experiments for practical applications. 2001 Elsevier Science Inc. All rights reserved.
1. Introduction
Induction heating consists in using electric energy dissipation to raise the temperature of a
metal workpiece in order to mechanically treat it. This process has numerous advantages over
other heating methods in the sense that the generated electric current is simpler to control than for
other processes and moreover the current characteristics (frequency, voltage, etc.) govern, in a
relatively understood way, the principal heating properties (temperature gradient, heating dura-
tion, phase transition, etc.).
We investigate in this paper a numerical method to control principal induction heating pro-
cesses through heating parameters such as shape of the inductor, frequency, current voltage, and
heating duration. For this, we use as an example a simple two-dimensional cartesian induction
heating model. It turns out that two main types of induction heating: uniform heating and metal
hardening lead to two dierent functionals to minimize.
Let us point out that the induction heating model has thoroughly been studied either from a
theoretical point of view [1] or from a numerical one [2].
The outline of the paper is as follows. In Section 2 we describe the state equations. Namely, we
derive a simple two-dimensional electromagnetic problem. Here we further assume that the in-
ductor is thin. This assumption leads, by an asymptotic analysis, to a well-conditioned problem
www.elsevier.nl/locate/apm
Applied Mathematical Modelling 25 (2001) 697712
*
Corresponding author. Tel.: +33-4-73-40-70-50; fax: +33-4-73-40-70-64.
E-mail addresses: obodart@ucfma.univ-bpclermont.fr (O. Bodart), a.boureau@voila.fr (A.-V. Boureau), touzani@ucfma.
univ-bpclermont.fr (R. Touzani).
0307-904X/01/$ - see front matter 2001 Elsevier Science Inc. All rights reserved.
PII: S 0 3 0 7 - 9 0 4 X ( 0 1 ) 0 0 0 0 7 - 5
and simplies the optimization of the inductor shape. The thermal problem here is a one-phase
model (no phase transition is handled). Section 3 presents the two types of optimal control
problems that will be solved. A diculty arises here in the choice of control parameters that
determine the inductor shape. Section 4 gives the optimality system to solve. These systems are
formally obtained by rst writing the Lagrangian of the problem and then using the classical
adjoint problem technique. Section 5 describes the numerical method used to solve the optimality
systems and gives the results of some numerical experiments. Finally, Section 6 gives some
concluding remarks about the possible generalizations of the model. In fact, a large part of the
simplifying assumptions introduced in the paper can be overcome and generalization of our
model can be simply obtained.
2. The induction heating model
We consider the following set-up: A cylindrical conductor that is assumed to be innite in the
direction of its generating line Oz. This cylinder is surrounded by a cylindrical inductor of a ring-
shaped section and in which ows an alternating time-harmonic current. The voltage of this
current is denoted by V . The geometrical conguration presented above allows us to seek a one-
component magnetic eld (the component along the z-axis).
Let X denote a section of this conductor in the xy-plane, with boundary C and let c denote a
closed smooth curve that generates the ring-shaped section of the inductor (cf. Fig. 1).
The induction heating model consists in a set of two equations: an electromagnetic problem
and a thermal problem.
2.1. The electromagnetic problem
The electromagnetic phenomenon is governed by Maxwell's equations using the low-frequency
hypothesis and a thin inductor approximation. Assuming that the current voltage v is a time
function of the form
Fig. 1. The conductor X and the inductor c.
698 O. Bodart et al. / Appl. Math. Modelling 25 (2001) 697712
v(t) = Re(e
ixt
V ); V C;
where x is the angular frequency. Due to the geometry of the conductor we seek a magnetic eld
of the form
H(x
1
; x
2
; x
3
; t) = Re(e
ixt
h(x
1
; x
2
))e
3
; (2:1)
where e
3
is the third vector of the canonical basis of R
3
. Maxwell's equations then become (see [2,3])
ixlh div(q\h) = 0 in X;
h = Const: on C; (2:2)
ixl
_
X
hdx ixlA
_

q
0
`
e
_
h
[C
= V :
Here, above l is the magnetic permeability of the free space, q, q
0
and e are, respectively, the
electric resistivities of the conductor X and the inductor c and the thickness of the inductor (in the
z-direction). In the following, we set a = q
0
=e. The parameters V , A and ` stand, respectively, for
the current voltage, the area delimited by C and c and the length of the curve c.
In [3] this model was rigorously justied by an asymptotic analysis. In particular, it is note-
worthy that in the case where the conductor X is not present, we obtain the well-known Kirchho
circuit equation. Another important fact to note about the above model is that the magnetic eld
depends on the inductor only by the parameters A and ` and not explicitly by its shape. Such a
feature simplies the setting of the electromagnetic problem but, as we shall see later, does not
simplify the optimization problem as well.
A variational formulation of Problem (2.2) is obtained by choosing the space
X = v H
1
(X); v
[C
= Const::
We obtain:
Find h X such that
ixl
_
X
hv
+
dx
_
X
q\h \v
+
dx (ixlA a`)h
[C
v
+
[C
= Vv
+
[C
\v X;
_
(2:3)
where v
+
is the complex conjugate of v. The setting of this problem can be simplied by a
transformation. Let h denote the solution of this problem and let h
0
= h
[C
C and

h := h=h
0
. We
obtain for

h the variational problem:
div(q\

h) ixl

h = 0 in X;

h = 1 on C; (2:4)
h
0
=
V
ixl(
_
X
hdx A) a`
:
The value h
0
can then be retrieved by setting v = 1 in Problem (2.3). We obtain
h
0
=
V
ixl((

h; 1) A) a`
: (2:5)
2.2. The thermal problem
The instantaneous power density can be expressed as q[\ H[
2
. According to (2.1) and
averaging on one period (0; 2p=x) we obtain as an averaged power density the term
O. Bodart et al. / Appl. Math. Modelling 25 (2001) 697712 699
x
2p
_
2p=x
0
q[\ Re(e
ixt
H)[
2
) dt =
q
2
[\h[
2
;
where [v[
2
= v
1
v
+
1
v
2
v
+
2
for a complex-valued vector eld v = (v
1
; v
2
).
Moreover, we shall consider the general case where the supplied heating power acts only during
a time interval (0; T
0
[ while the current is switched o during the remaining time interval
(T
0
; T
0
T
1
[. That is, the right-hand side of the heat equation will be a discontinuous function
f (h
0
;

h) :=
q
2
[h
0
[
2
[\

h[
2
in X (0; T
0
[;
0 in X (T
0
; T
0
T
1
[:
_
_
_
This means that the parameter T
0
is the heating time duration while the parameter T
1
is a tem-
perature smoothing time.
Assuming that the conductor X is thermally insulated, we have the heat equation:
C
oh
ot
div(a\h) = f (h
0
;

h) in X (0; T
0
T
1
[;
a
oh
on
= 0 on C (0; T
0
T
1
[; (2:6)
h = h
0
in X; t = 0;
where C is the specic heat, a is the thermal conductivity, h
0
is the initial temperature and oh=on is
the normal derivative of h on C.
Clearly, since the values of T
0
, T
1
will be the control parameters we shall consider a change of
time variable in the previous problem. Namely, we dene the transformation:
u(t) =
t
T
0
; 0 6t 6T
0
;
1
t T
0
T
1
; T
0
6t 6T
0
T
1
:
_

_
Denoting by

h the function

h(x; u(t)) := h(x; t); x X; t (0; T


1
T
2
):
Problem (2.6) becomes
C
o

h
o

t
T
0
div(a\

h) =
qT
0
2
[h
0
[
2
[\

h[
2
in X (0; 1[;
C
o

h
o

t
T
1
div(a\

h) = 0 in X (1; 2[;
a
o

h
on
= 0 on C (0; 2[;

h = h
0
in X;

t = 0;
(2:7)
where

t = u(t). In the following we shall omit the ``hat'' over h in order to simplify the notations.
Problems (2.4) and (2.7) constitute the state problem. It is clear that we have chosen a quite
simple model since Problems (2.4) and (2.7) are not coupled. A more realistic version consists in
taking the parameters C, a and q as depending on the temperature h and even, for a nonmagnetic
material, l may depend on h and h. This would make more complicated the setting of the problem
but does not fundamentally change the purpose of the present work. For this reason, we have
omitted to keep the whole generality of the problem.
700 O. Bodart et al. / Appl. Math. Modelling 25 (2001) 697712
3. The optimal control problem
We are now concerned with the optimal control of the induction heating process. More pre-
cisely, given an objective to reach (quality of the heating, a temperature to reach, etc.) one seeks
the best set of parameters of problem (2.4)(2.7) to realize the objective. Of course, to each type of
induction heating process corresponds at least a particular choice of the objective (or cost)
function.
As we have mentioned earlier, one can distinguish at least two types of induction heating
processes: uniform heating and metal hardening.
Uniform heating consists in heating the whole workpiece in order to ``treat'' it. An ideal sit-
uation would consist in obtaining a constant temperature at a desired nal time. But this can-
not be reasonably expected. At least one could hope of minimizing the gradient of the
temperature.
Metal hardening consists in raising the temperature until a critical value is reached then cool-
ing the metal according to a cooling curve in order to make irreversible the phase transfor-
mation is thus realized. Such a problem requires a much more complicated model than the
simple linear heat equation. For more description of the modeling of this phenomenon and
its numerical simulation and mathematical analysis, the reader can consult for instance [4]
or [5]. Here, as a simplied version of the problem we shall try to prescribe a given isothermal
line to be reached, in a sense to determine, at a given nal time.
3.1. Parameterization of the inductor
Before considering the set of control parameters particular attention is to be given to the in-
ductor shape. Here, the main diculty is that although the magnetic eld, and consequently the
temperature, depend only on the inductor by the parameters A and ` (respectively, the area be-
tween C and c and the length of c), giving two positive real numbers A and ` does not ensure the
existence of at least one curve c having these characteristics and even if this was possible, it does
not give any method to determine c. In other words, there is a geometrical constraint between A, `
and X. But this constraint is hard to obtain explicitly.
A remedy to this situation is to consider as curves c all smooth curves that are obtained from C
as deformations along external normal to C. The advantage of such a method is that it gives
explicitly a parameterization of the inductor curve c but the drawback is that the set of parameters
is innite. This is however not so dramatic since, as far as numerical approximation is considered,
the two parameters A and ` are replaced by the number of nodes on C.
Let us assume that the closed curve dening C is smooth enough and let
g
0
: [0; 1[ R
2
denote a parameterization of this curve such that
g
0
(0) = g
0
(1); g
/
0
(0) = g
/
0
(1):
For a positive smooth function s : [0; 1[ R we consider as a parameterization of the curve c:
g
s
: t [0; 1[ g
s
(t) := g
0
(t) s(t)n(t) R
2
;
where n(t) is the outward unit normal to C at point t. Of course one has to impose
s(0) = s(1); s
/
(0) = s
/
(1);
s(t) P0; t [0; 1[:
O. Bodart et al. / Appl. Math. Modelling 25 (2001) 697712 701
According to this denition, one can obtain by simple integral calculus the values of A and ` for a
function s. We have from the GreenRiemann formula:
A(s) =
_
1
0
s(t)g
/
0
(t)
_

1
2
s
2
(t)j(t)
_
dt; (3:1)
`(s) =
_
1
0
([g
/
0
(t)[
_
s(t)j(t))
2
s
/
(t)
2
_
dt; (3:2)
where j(t) := [n
/
(t)[ is the curvature of C. Note that other denitions of c could be possible (e.g.,
replacing the normal by a radial vector), but the calculation of A(s) and `(s) could be much more
complicated.
3.2. Setting of the optimal control problem
The optimal control problems that under consideration will be dened through the cost
function to minimize and the set of admissible controls. The control parameters are x, V , T
0
, T
1
and s.
1. In the case of uniform heating, we seek a temperature eld as close as possible to a given
maximum temperature h
M
at the nal time T
1
. Moreover, we want to attain this goal with a
minimal power supply. The cost function will thus be of the form:
J(x; V ; T
0
; T
1
; s) :=
1
2[X[
_
X
1
_

h(x)
[t=2
h
M
_2
dx
1
2
V
2
: (3:3)
2. In the case of metal hardening, we would like to prescribe a closed isothermal line that
separates the treated zone from the nontreated one at the nal time. Here also a minimal power
supply is sought. Then, if K denotes a closed curve that will denote the contour line with tem-
perature h
M
, we dene
J(x; V ; T
0
; T
1
; s) :=
1
2[X[
_
K
1
_

h(x)
[t=2
h
M
_2
dr
1
2
V
2
: (3:4)
Note that in both cases, the rst term in the functional is nondimensional in order to get a data
independent functional. The second term cannot be unfortunately scaled.
The set of admissible controls is dened by
U
ad
:= (x; V ; T
0
; T
1
; s) R

H
1
per
(0; 1); `(s) 6`
M
: (3:5)
Here,
H
1
per
(0; 1) := s H
1
(0; 1); s P0; s(0) = s(1):
One may note that all the control variables are nonnegative and that we have a nonlinear con-
straint on the function `. In fact, considering the expressions (3.1) and (3.2), we can easily prove
that if the function `(s) is bounded from above then s is bounded and consequently A(s) is also
bounded from above. The choice of the set H
1
per
(0; 1) for s is clearly justied.
The optimal control problem is nally then:
Find (x; V ; T
0
; T
1
; s) U
ad
such that
J(x; V ; T
0
; T
1
; s) = min
(x;V ;T
0
;T
1
;s)U
ad
J(x; V ; T
0
; T
1
; s):
_
_
_
(3:6)
702 O. Bodart et al. / Appl. Math. Modelling 25 (2001) 697712
3.3. The optimality system
We now turn to the derivation of the optimality system. This will be done formally. We shall
dene a Lagrangian associated to the functional J. Assuming a convexity hypothesis in a
neighborhood of an optimum, the optimality system can then be obtained by dierentiating in the
direction of each variable.
The equality constraint dening the set U
ad
is approximated by a penalty method. The opti-
mality equations are consequently dened by cancelling the derivatives of the Lagrangian in the
direction of each of the variables and using appropriately Green's formula.
Let us denote by

k, k
0
, p the adjoint variables (i.e., the Lagrange multipliers) associated to the
state variables

h, h
0
, and h, respectively. Let us, for the sake of conciseness, denote by (; ) the
scalar product in L
2
(X), i.e.
(u; v) :=
_
X
u(x)v
+
(x) dx:
The penalized Lagrangian can be written as
L(x; V ; T
0
; T
1
; s;

h; h
0
; h;

k; k
0
; p)
:= J(x; V ; T
0
; T
1
; s; h) r 1
_ _

`(s)
`
M
_

_
2

_
2
0
C
oh
ot
; p
_ _
dt
T
0
2
_
1
0
q[\

h[
2
[h
0
[
2
; p) dt
T
0
_
1
0
(a\h; \p) dt T
1
_
2
1
(a\h; \p) dt
Re(ixl(

h;

k) (q\

h; \

k))
Re(h
0
k
+
0
(ixl(

h; 1) ixlA(s) q
0
`(s)) V k
+
0
);
where r is a suciently large positive number and g

denotes the positive part of the function g.


We now calculate the partial derivatives of Lwith respect to the dierent variables in order to
obtain the optimality system.
1. Let
p =
_
2
0
p(t) dt:
The derivative with respect to

h gives the adjoint equation for

k. We have indeed
o~
h
L(x; V ; s;

h; h
0
; h;

k; k
0
; p) d

h
:= [h
0
[
2
(qp; Re(\

h \d

h
+
)) Re(ixl(d

h;

k) (q\d

h; \

k)) Re(ixlh
0
k
+
0
(d

h; 1)):
(3:7)
Therefore, cancelling this derivative gives the problem:
Find

k H
1
0
(X) such that
ixl(

k; v) (q\

k; \v) = [h
0
[
2
(qp\

h; \v) ixlh
+
0
k
0
(1; v) \v H
1
0
(X):
_
(3:8)
O. Bodart et al. / Appl. Math. Modelling 25 (2001) 697712 703
2. To obtain a relationship dening k
0
we dierentiate L with respect to h
0
:
o
h
0
L(x; V ; s;

h; h
0
; h;

k; k
0
; p) = h
0
(qp; [\

h[
2
) Re((ixl((

h; 1) A(s)) a`(s))k
+
0
): (3:9)
Cancelling this derivative yields
k
0
=
h
0
(qp; [\

h[
2
)
a`(s) ixl((

h; 1) A(s))
: (3:10)
3. Let us now dierentiate L with respect to h to get an equation for its adjoint variable p:
o
h
L(x; V ; s;

h; h
0
; h;

k; k
0
; p)dh
= o
h
J(x; V ; T
0
; T
1
; s; h) dh
_
2
0
C
odh
ot
; p
_ _
dt
T
0
_
1
0
(a\p; \dh) dt T
1
_
2
1
(a\p; \dh) dt: (3:11)
Whence the backward parabolic problem:
C
op
ot
T
0
div(a\p) = 0 in X [0; 1);
C
op
ot
T
1
div(a\p) = 0 in X [1; 2);
op
on
= 0 on oX [0; 2);
p(2) =
1
C
o
h
J(x; V ; T
0
; T
1
; s; h)oh in X:
(3:12)
4. The derivative of L with respect to k
0
is given by
o
k
0
L(x; V ; s;

h; h
0
; h;

k; k
0
; p) = h
+
0
(ixl((

h; 1) A(s)) a`(s)) V :
From this we obtain again relationship (2.5) with A and ` depending on s:
h
0
=
V
(ixl((

h; 1) A(s)) a`(s))
: (3:13)
5. To obtain the equation for h we dierentiate L with respect to p. We retrieve the heat
equation (2.6).
6. Finally, we obtain the optimality conditions by dierentiating Lwith respect to u. We have:
o
x
L(x; V ; T
0
; T
1
; s;

h; h
0
; h;

k; k
0
; p) = Re(il((

h;

k) h
0
k
+
0
((

h; 1) A(s)))); (3:14)
o
V
L(x; V ; T
0
; T
1
; s;

h; h
0
; h;

k; k
0
; p) = o
V
J(x; V ; T
0
; T
1
; s; h) Rek
0
; (3:15)
o
T
0
L(x; V ; T
0
; T
1
; s;

h; h
0
; h;

k; k
0
; p) =
_
1
0
(a\h; \p) dx
1
2
[h
0
[
2
_
1
0
(q[\

h[
2
; p) dx; (3:16)
o
T
1
L(x; V ; T
0
; T
1
; s;

h; h
0
; h;

k; k
0
; p) =
_
2
1
(a\h; \p) dx; (3:17)
o
s
L(x; V ; T
0
; T
1
; s;

h; h
0
; h;

k; k
0
; p) =
r`
/
(s)
`
M
1
_

`(s)
`
M
_

Re(h
0
k
+
0
(ixlA
/
(s) a`
/
(s))):
(3:18)
704 O. Bodart et al. / Appl. Math. Modelling 25 (2001) 697712
4. Numerical solution
Let us now consider the numerical solution of the above problem. Clearly, we have to use an
optimization algorithm. In the following paragraph, we shall describe the minimization algorithm
as well as the ow chart of the numerical code that solves the problem.
4.1. Optimization procedure
Here we are using the BFGS algorithm in a limited memory version as implemented by Byrd
et al., [6]. In order to use this software, one has to provide the cost function J and its partial
derivatives with respect to each variable. Note here that, thanks to the penalization, we have
an optimization problem over the set
U
+
ad
:= (x; V ; T
0
; T
1
; s) R

H
1
per
(0; 1): (4:1)
The above set exhibits only simple bound constraints that can be taken into account by a pro-
jection method.
The ow chart of our optimization program is then:
1. Initialize control parameters (x
0
; V
0
; T
0
0
; T
0
1
; s
0
);
2. set k = 0;
3. Calculate the magnetic eld (

h; h
0
) by solving problems (2.4)(3.13);
4. Calculate the temperature h by solving (2.6);
5. Calculate p by solving the backward parabolic problem (3.12);
6. Calculate (

k; k
0
) by using identity (3.10), then solving problem (3.8);
7. Evaluate J as given by (3.4) or (3.6);
8. Calculate derivatives of J by relationships (3.14)(3.18);
9. Update control parameters by calculating (x
k1
; V
k1
; T
k1
0
; T
k1
1
; s
k1
) by BFGS process or
any other descent algorithm;
10. Project control parameters on the set U
+
ad
;
11. If Convergence then stop; otherwise set k := k 1 and go to (3).
4.2. Discretization of the equations
The involved partial dierential equations require using a numerical discretization method to
solve them. Here we use a nite element method to discretize in space the equations. After a
partitioning of the domain X into triangles, we approximate Eqs. (2.4) and (2.6) by using a P
1
-
triangular nite element method. The time discretization uses the implicit Euler scheme.
4.3. Numerical calculation of the derivatives
Numerical calculation of the cost function and its derivatives constitutes, in optimization
problems, an important issue. Here, the most challenging diculty is the calculation of quantities
relative to the inductor shape.
The rst diculty arises when one wants to localize the points in the inductor. Namely, for a
polygonal domain, the normal to C exhibits a discontinuity in the angles and one has to smooth
this function. For this end, we use a method that was primarily used by Engelman et al., [7] to
prescribe slip boundary conditions in uid ow problems. In their paper, simple mass conser-
vation properties lead to the following denition of a smoothed normal vector. Namely, at each
node k on the boundary of X we dene a unit normal vector by
O. Bodart et al. / Appl. Math. Modelling 25 (2001) 697712 705
n
k
i
=
_
X
(ou
k
=ox
i
) dx
_
X
(ou
k
=ox
1
) dx
_ _
2

_
X
(ou
k
=ox
2
) dx
_ _
2
_ _
1=2
; i = 1; 2;
where u
k
is the nite element basis function associated to node k. It can be easily checked that if k
does not exhibit a corner then the above denition coincides with the outward unit normal to k
and that if k is in a corner then this denition corresponds to a weighted averaging of normals to
neighbor nodes to k. The function n(t) is therefore approximated by interpolation by a
continuous function that is ane in each segment constituting the boundary C. Its derivative
and then the curvature is consequently a piecewise constant function. Denoting still by n(t)
(resp. j(t)) the obtained approximation of the normal (resp. the curvature), we calculate an ap-
proximation of the area A(s) by numerical integration thus obtaining
A(s) ~

N1
j=1
d
2
(s
j
_
s
j1
)[g
/
j
[
d
6
(s
2
j
s
j
s
j1
s
2
j1
)j
j
_
;
where C is considered as union of N segments of equal length d and s
j
, g
/
j
, j
j
are values (or
approximations) of s, g
/
, j at the jth vertex of C. Likewise, an approximation of `(s) is obtained
by the formula
`(s) ~
d
2

N1
j=1
([g
/
j
s
j
n
/
j
s
/
j
n
j
[ [g
/
j
s
j1
n
/
j
s
/
j
n
j1
[):
Computing approximations of the derivatives of A and ` with respect to s is much more fastidious.
Denoting by e
k
the kth vector of the canonical basis of R
N
and by d
jk
the Kronecker delta, we
nally obtain
A
/
(s)e
k
~

N1
j=1
d
2
[g
/
j
[(d
jk
_
d
j1;k
)
d
6
j
j
(2s
j
d
jk
s
j
d
j1;k
2s
j1
d
j1;i
)
_
:
In an analogous way we have
`
/
(s)e
k
~

N1
j=1
dg
/
j
s
j
n
/
j
(1=d)(s
j1
s
j
)n
j
; n
/
j
d
jk
(1=d)(d
j1;k
d
jk
))
4[g
/
j
s
/
j
n
/
j

1
d
(s
j1
s
j
)n
j
[

N1
j=1
dg
/
j
s
j1
n
/
j
(1=d)(s
j1
s
j
)n
j1
; n
/
j
d
j1;k
(1=d)(d
j1;k
d
jk
))
4[g
/
j
s
/
j
n
/
j1
(1=d)(s
j1
s
j
)n
j1
[
5. Numerical experiments
In order to validate the model and numerical methods described above we have tested them on
realistic congurations. This is made through the following signicant examples.
5.1. ``Uniform'' heating of a square-shaped conductor
We consider a square conductor (0; 0:2) (0; 0:2) made of Stainless Steel. The cost function to
minimize is given by (3.3). Material data are those of Ref. [2]. Namely,
706 O. Bodart et al. / Appl. Math. Modelling 25 (2001) 697712
l = 1:2566 10
6
; q = 7:1265 10
7
X m; q
0
= 1:6778 10
8
X m;
k = 15:342 W m
1
K
1
; C
p
= 3:0427 10
6
J K
1
m
3
:
The initial temperature is h
0
= 293 K. Initial values for control parameters were chosen with the
following values:
x = 10000 Hz; V = 20 V; T
0
= T
1
= 30 s; s(t) = 0:01 m \t:
Fig. 2 shows the nite element mesh. According to the small skin depth the mesh was chosen to be
ne near the boundary.
The desired temperature was chosen equal to h
M
= 1000 K.
The optimization algorithm converges in ve iterations thus giving as control parameters the
values
x = 10000 Hz; V = 20 V; T
0
= T
1
= 30 s; s
i
= 0:0039456 m \i:
As can be noticed, the initial constant value of the function s was kept constant during the it-
erations. The following gures give the values of cost function and the norm of the projected
Fig. 2. Finite element mesh.
O. Bodart et al. / Appl. Math. Modelling 25 (2001) 697712 707
gradient as a function of the iterations. Furthermore, the optimized solution involves only a
change in the distance of the inductor from the domain (cf. Fig. 3).
5.2. ``Uniform'' heating of a nonconvex domain
The following example tests the eciency of the determination of the inductor shape even for
nonconvex domains. We consider here the domain given in Fig. 4 (with its nite element mesh)
and the computed inductor shape.
Fig. 3. Cost function and gradient iteration history.
708 O. Bodart et al. / Appl. Math. Modelling 25 (2001) 697712
Material properties were taken to be the same as in the previous example. The initial tem-
perature here is also h
0
= 293 K. Initial values for control parameters were chosen as
x = 1000 Hz; V = 50 V; T
0
= T
1
= 30 s; s(t) = 0:004 m \t:
The desired temperature is h
M
= 900 K.
The optimization algorithm converges in nine iterations thus giving as control parameters the
values
x = 998 Hz; V = 32:8 V; T
0
= 17:155 s; T
1
= 31:085 s; s
i
= 0:005 m \i:
Here also the obtained inductor exhibits a constant function s. The following gures give
the values of cost function and the norm of the projected gradient as a function of the
iterations.
This test shows the eciency of the inductor parameterization. It is to be noted however, that
the model can handle only an inductor which is close enough to the conductor in the vicinity of
the entering angle. This can be imposed by adding a constraint on the function s near this
corner.
5.3. ``Hardening'' of a square conductor
In this third example we investigate the case of the second choice of the cost function (3.4). We
use the same material as in the previous examples. The domain is a square of edge length 0.02 m.
The initial control parameters are:
x = 1000 Hz; V = 50 V; s(t) = 0:004 m \t:
Note that here T
0
and T
1
are not the control parameters. We have chosen T
0
= 30 and T
1
= 0.
Fig. 4. Finite element mesh of a nonconvex domain.
O. Bodart et al. / Appl. Math. Modelling 25 (2001) 697712 709
The initial temperature is here h
0
= 293 K. Initial values for control parameters were chosen as
x = 1000 Hz; V = 50 V s; s(t) = 0:004 m \t:
The desired temperature on a curve K was chosen as equal to h
M
= 1123 K.
The optimization algorithm converges in ve iterations (Fig. 5) thus giving as control pa-
rameters the values
Fig. 5. Cost function and gradient iteration history.
710 O. Bodart et al. / Appl. Math. Modelling 25 (2001) 697712
x = 62832 Hz; V = 10 V; s
i
= 0:005 m \i:
Here also the initial constant value of the function s was kept constant during the iterations. Fig. 6
gives the domain X as well as the curve K and the isothermal line corresponding to this curve. It is
noteworthy that even though a temperature contour cannot have angles, thanks to smoothing
properties of the heat equation, the obtained isothermal line is as close as possible to the curve K.
6. Concluding remarks
To summarize, we have developed an optimization procedure for induction heating processes.
The goal is to control the quality of heating and to prescribe a temperature to reach either in the
whole domain or on a prescribed closed curve. The induction heating model concerns simple two-
dimensional cartesian geometries and the electromagnetic and thermal problems are assumed to
be uncoupled. These restrictions are however not essential in the model.
Fig. 6. The curve K and the obtained isothermal line.
O. Bodart et al. / Appl. Math. Modelling 25 (2001) 697712 711
1. Coupling between electromagnetic and thermal equation can be taken into account by as-
suming that electric resistivity q and magnetic permeability l depend on the temperature eld.
This feature can be implemented in the model without any extra diculty.
2. The heat equation is provided with a homogeneous Neumann boundary condition thus
assuming thermal insulation of the boundary. Radiation or convection boundary conditions can
be supplied to the model as in [2] without aecting the validity of the model. Let us mention here
that minimizing the losses due to radiation and/or convection can be added in the cost function.
3. The two-dimensional character of the equations is important but not fundamental. In [8], a
three-dimensional axisymmetric model is investigated where the shape optimization of the in-
ductor exhibits, as one can expect, more serious diculties.
4. The choice of the two cost functions presented in this paper may be not the best possible.
Other choices could be made. In particular, rather than prescribing a mean temperature one could
involve a maximal temperature. But this can reveal an extra diculty since constraints on the
state variables are more dicult to handle than those on control parameters.
6. It is clear that the so-called hardening case in this paper does not describe the right model for
such a complex phenomenon. We have simply used a linear heat equation. We claim however that
the model described in [2] can be implemented here to give more realistic results.
References
[1] S. Clain, Modelisation mathematique et analyse numerique d'un probleme de chauage par induction, Ph.D. Thesis,
Ecole Polytechnique Federale de Lausanne, Switzerland, 1994.
[2] S. Clain, J. Rappaz, M. Swierkosz, R. Touzani, Numerical modeling of induction heating for two-dimensional
geometries, M
3
AS 3 (6) (1993) 805822.
[3] R. Touzani, Analysis of an eddy current problem involving a thin inductor, Comput. Meth. Appl. Mech. Eng. 131
(1996) 233240.
[4] A. Jacot, M. Swierkosz, J. Rappaz, M. Rappaz, D. Mari, Modelling of Electromagnetic heating, cooling and phase
transformations during surface hardening of steels, Mecamat'95, J. Phys. IV (C1) (1996) 203213.
[5] D. H omberg, A mathematical model for the phase transitions in eutectoid carbon steel, IMA J. Appl. Math. 54
(1995) 3157.
[6] R.H. Byrd, P. Lu, J. Nocedal, C. Zhu, A limited memory algorithm for bound constrained optimization, SIAM J.
Scientic Comput. 16 (5) (1995) 11901208.
[7] M.S. Engelman, R.L. Sani, P.M. Gresho, The implementation of normal and/or tangential boundary conditions in
nite element codes for incompressible uid ow, Int. J. Numer. Meth. Fluids 2 (1982) 225238.
[8] O. Bodart, R. Touzani, Optimal control problems in electromagnetic processess, SIAM Annual Meeting, Toronto,
Canada, 1998.
712 O. Bodart et al. / Appl. Math. Modelling 25 (2001) 697712

You might also like