You are on page 1of 4

IEEE TRANSACTIONS ON MAGNETICS, VOL. 36, NO.

4, JULY 2000 809

Numerical Calculation of Nonlinear Transient Field


Problems with the Newton–Raphson Method
Silvia Drobny and Thomas Weiland

Abstract—The modeling and numerical simulation of transient (MGE). The transformation is based on a dual orthogonal grid
electromagnetic field problems with ferromagnetic materials is de-
scribed in the context of the Finite Integration Technique (FIT). system using voltages along grid lines and fluxes
This paper presents a Newton–Raphson method adapted to FIT
which is faster than the Successive Approximation technique. Due over areas as degrees of freedom:
to the nonlinear material behavior these algorithms are not al-
ways stable. In order to improve convergence, a dynamic relaxation (1)
process and a suitable interpolation of the magnetization curve are
presented. Results of these iterative schemes are compared with
measurements of typical benchmark problems. (2)
Index Terms—Quasistatic fields, eddy currents, nonlinear prob-
lems, ferromagnetic materials. (3)
(4)
I. INTRODUCTION
(5)

A PPLICATIONS of the FIT discretization process to the nu-


merical simulation of transient electromagnetic field prob-
lems with ferromagnetic materials were first described in [1].
(6)

For the situation of a slowly-varying field process, the qua- (7)


sistatic approach yields a differential algebraic system which
uses the primitive of the electric grid voltage as primary vari- In this formulation denotes the impressed source current
able. Since the permeability depends on the local flux density vector and refers to the space charge density.
resulting in nonlinearity, a nonlinear equation system has to be The correspondence between both grids is given by (5)–(7)
solved at each time step. with the matrices and containing averaged ma-
For this the Newton–Raphson method offers a faster conver- terial properties and mesh dimensions.
gence behavior than the Successive Approximation technique The so called curl and source matrices describe
presented in [2]. Because in saturated regions stability is not the topology of the grids and contain only entries in .
guaranteed, a dynamic relaxation process is introduced which It can be shown that due to the duality of the two grids the fol-
takes the nonlinear material behavior into account and reduces lowing property holds:
the number of nonlinear cycles. Convergence problems are also
(8)
avoided by considering the special form of the magnetization
curve and choosing an appropriate interpolation scheme. For This yields the correspondence to the analytical relations
this the choice of the employed scheme for computing the new
permeability has an authoritative influence on the convergence (9)
of the iterative algorithm. (10)
Finally the Successive Approximation technique and the
Newton–Raphson method are compared by the results of two which ensure the divergence conservation on the grid system
transient problems of the TEAM Workshop. and unique numerical solutions among many other constitutive
relations [4].
II. FINITE INTEGRATION TECHNIQUE
III. TIME-DOMAIN FORMULATION
The Finite Integration Technique [3] transforms the contin-
uous Maxwell Equations in their integral form into a set of con- A suitable modeling approach to slowly-varying problems
sistent matrix equations, yielding the Maxwell-Grid-Equations consists in the omission of the displacement currents in the set
of Maxwell-Grid-Equations. Using the primitive of the elec-
Manuscript received October 25, 1999; revised January 17, 2000.
The authors are with the Fachbereich 18 Elektrotechnik und Informa-
tric grid voltage vector as primary vector of unknowns yields
tionstechnik, Fachgebiet Theorie Elektromagnetischer Felder, Technische a quasistatic formulation
Universität Darmstadt, Schloßgartenstr. 8, D-64289 Darmstadt, Germany
(e-mail: {drobny; weiland}@temf.tu-darmstadt.de).
Publisher Item Identifier S 0018-9464(00)06876-X. (11)

0018–9464/00$10.00 © 2000 IEEE


810 IEEE TRANSACTIONS ON MAGNETICS, VOL. 36, NO. 4, JULY 2000

which requires to solve a differential-algebraic system of equa- and results from solving the linear equa-
tions (DAE) of index 1 [5]. The time-integration of this system tion system
is performed using suitable implicit time-marching schemes,
known to be unconditionally stable [6].
Since the iteration matrix is symmetric and positive semidef- (13)
inite, the solution of the linear systems is evaluated using effi-
ciently implemented preconditioned conjugate gradient (PCG)
methods which are robust even for non gauged systems. They with the relaxation factor and the index denoting
also have low memory requirements and feature a superlinear the -th iteration step [7].
convergence behavior.
D. Magnetization Curves
IV. NONLINEAR PROBLEM The magnetic material is described by discrete values for
and which are interpolated with straight lines resulting in
In case of ferromagnetic materials with a nonlinear magneti-
a -curve which is used in the nonlinear cycles. The cal-
zation curve the material matrix depends on the unknown
culation of the numerical derivatives required for the Jacobian
vector potential resulting in nonlinearity of (11). The relation matrix in (13) is performed using a -curve. The differen-
between material values and field intensities is normally given tial permeability which corresponds to the iteration matrix
by discrete characteristics and not representable by simple oper- in (13) is updated by using the BH- or HB-scheme as described
ators which could be incorporated in the system. Thus for each above.
time step a nonlinear system of equations of the form The choice of the employed scheme for computing the new
permeability has an authoritative influence on the convergence
(12) of the iterative algorithm due to the form of the interpolated
magnetization curve of the nonlinear material [8]. In case that
has to be solved by a sequence of linear problems with the the resulting magnetization curve is monotonous and convex
system matrix and the right hand side vector . This in- the permeability decreases monotonous. In this situation the
creases the number of linear algebraic systems to be solved for BH-scheme is stable and faster than the HB-scheme. However,
the whole problem. in several practical cases the magnetization curve is s-formed
and concave for low field values such that the permeability first
A. Successive Approximation Technique increases before decreasing. This may lead to a violation of
A stable, but in some cases slow method is the Successive physical relations in the flat part of the concave curve due to os-
Approximation technique [2] which solves the nonlinear equa- cillations: too small values of the permeability in one cycle can
tion by splitting it up in a sequence of linear problems for each lead to much too high values in the next cycle and vice versa.
of which a constant material distribution is assumed. Thus it might take very long to obtain a correct value. In conse-
quence the stability of the above BH-scheme is no more ensured
B. Nonlinear Cycles and it is convenient to employ the HB-scheme.
For this situation the calculation of the new time step solu-
E. Relaxation Process
tion can be performed by two different schemes. First the mag-
The use of a relaxation parameter is a common technique to
netic facet flux is calculated by . Then in the
either achieve convergence or to improve the convergence speed
BH-scheme the value for the dual magnetic grid voltage is of an iterative scheme.
obtained from the magnetization curve. In contrast to this the A well considered damping of the low frequency oscillations
introduced by the succession of linear cycles is very important
HB-scheme first multiplies the magnetic flux with the ma-
for an effective nonlinear solution scheme. Additionally conver-
terial matrix to yield the magnetic voltage . Then the gence problems may occur in the concave part of an s-formed
magnetization curve as well as in saturated regions. They can
new value for the magnetic flux is obtained from the magne-
also be reduced by weighting the new value of the permeability
tization curve. After this the value for the new permeability is
with the old one with a factor less then one.
derived from the secant vector of the magnetic flux and the Normally it is not possible to predict the optimal relaxation
parameter for the -th nonlinear cycle, because this param-
magnetic voltage .
eter depends on local field values, the actual time step and the
This cycle is repeated until the variation of the permeability
cycle course. The iteration is started with a maximal relaxation
between two steps becomes sufficiently small or up to a max-
parameter to get with big steps near to the searched mate-
imum of 20 cycles per time step.
rial distribution. In every cycle a value is calculated which
C. Newton–Raphson Method indicates the maximal deviation of the permeability in the region
of nonlinear material. Ideally this indicator yields continuous
The Newton–Raphson method offers a faster convergence be- decreasing values. The parameter is reduced if this indicator
havior by solving the nonlinear equations using a Taylor-ex- is greater than in the previous cycle. After a certain period of
pansion. The iteration is carried out by calculating convergence the parameter is increased again by . Hence,
DROBNY AND WEILAND: NUMERICAL CALCULATION OF NONLINEAR TRANSIENT FIELD PROBLEMS 811

Fig. 1. The discretized geometry of the TEAM Workshop problem 10 Fig. 3. The discretized geometry of model A of the TEAM Workshop
consists of an exciting coil placed between two steel channels and a steel problem 21 consists of two steel plates, one of them with a hole in the middle,
plate inserted between the channels leaving two air gaps. It is required to find whereas model B consists of a single steel plate with the same measurements.
the time functions of the average magnetic flux density over the three cross The plates are set near to two air-core exciting coils which are driven with
sections S1, S2 and S3. currents of opposite sign. Measured values are available for the x-component
of the magnetic flux density along four paths in the air gap between the steel
plates and the coils.

Fig. 2. The measured values for the average magnetic flux density of the
TEAM Workshop problem 10 are compared with the results of the nonlinear Fig. 4. The curves compare the computed values for the x-component of the
transient calculation. magnetic flux density with the measurements along path I and II of model B of
the TEAM Workshop problem 21.

the adaption of the relaxation parameter to the field problem is


by a fixed until a minimal is reached. Hence, is
easily reached.
adapted to the specific field problem, the stage of the -itera-
Herein a local strategy is included to circumvent that different
tion and the time step .
local regions are iterated with the same relaxation parameter. In
critical sections the so-called -factor
V. NUMERICAL RESULTS

(14) Two large 3D nonlinear eddy current problems have been


solved using the presented solution methods and the Crank-
Nicolson scheme for time discretization.
which gives the relation of the relative changes of the perme-
ability and the field strength, is greater than one. This situation
may occur in the flat part of an s-formed magnetization curve A. TEAM Problem 10
and in saturated regions. It is useful to reduce the local relax- The discretized geometry of the TEAM Workshop problem
ation parameter in these sections. The actual relaxation param- 10 is shown in Fig. 1. Since the structure has three symmetry
eter is now calculated dependent on the local field values, the planes only an eighth is computed with about 250.000 un-
-factor, the value of and a global , which serves as knowns. The magnetization curve is monotonous and convex
a guide value. In order to reach a maximal speed of convergence so that the BH-scheme is used. The results for the nonlinear
the starting parameter is set on a high value . If the indi- transient calculation are shown in Fig. 2 along with the
cator shows convergence difficulties, the parameter is reduced measurements.
812 IEEE TRANSACTIONS ON MAGNETICS, VOL. 36, NO. 4, JULY 2000

The HB-scheme is employed because of the s-formed magneti-


zation curve. The measured values are compared in Figs. 4 and
5 with the results of the nonlinear transient calculation. The dif-
ference between the curves may denote that a hysteresis model
for the steel plates might be needed in addition.

C. Comparison of Nonlinear Schemes


Table I shows the number of necessary nonlinear cycles per
time step for both nonlinear methods over the PCG solution ac-
curacy for each simulated problem. The numerical experiments
were carried out in double precision to prevent round off errors
and the value was set to one.
It can be seen, that in case of saturated materials as in problem
Fig. 5. The computed values for the x-component of the magnetic flux density 10, the number of nonlinear cycles increases, when the required
are compared with the measurements along path III and IV of model A of the accuracy for the linear solution step is decreased. Additionally
TEAM Workshop problem 21. for all examples the number of nonlinear cycles increases when
the time step length is increased and especially in this case the
TABLE I Newton–Raphson method offers a better convergence behavior.
COMPARISON BETWEEN SUCCESSIVE APPROXIMATION TECHNIQUE (SA)
AND NEWTON–RAPHSON METHOD (NR)
VI. CONCLUSION
This paper investigates numerical methods for nonlinear
transient calculations with a time domain formulation based on
the FI-Technique. Two efficient methods have been presented
for the solution of the nonlinear algebraic equations arising
from time discretization. Most of the convergence difficulties
can be avoided by choosing appropriate procedures and relax-
ation parameters. Hence, the computed results for the TEAM
Workshop problems 10 and 21 are in good agreement with the
measurements.

REFERENCES
[1] M. Clemens, S. Drobny, and T. Weiland et al., “Time-integration of
slowly-varying electromagnetic field problems using the finite integra-
tion technique,” in Proceedings ENUMATH97, Bock et al., Eds., Singa-
pore, 1998, pp. 246–253.
[2] S. Drobny and T. Weiland, “Iterative algorithms for nonlinear transient
electromagnetic field calculation,” presented at the Proceedings
ISEM99, Pavia, Italy, 1999.
[3] T. Weiland, “Time domain electromagnetic field computation with fi-
nite difference methods,” International Journal of Numerical Modeling:
Electronic Networks, Devices and Fields, vol. 9, pp. 295–319, 1996.
[4] , “On the numerical solution of Maxwell’s equations and ap-
plications in accelerator physics,” Particle Accelerators, vol. 15, pp.
245–291, 1984.
[5] M. Clemens and T. Weiland, “Numerical algorithms for the FDiTD and
FDFD simulation of slowly varying electromagnetic fields,” Interna-
tional Journal of Numerical Modeling: Electronic Networks, Devices
and Fields, vol. 12, no. 1/2, pp. 3–22, 1999.
[6] K. E. Brenan, S. L. Campbell, and L. R. Petzold, “Numerical solution of
initial-value problems in differential-algebraic equations,” in Classics in
Applied Mathematics 14. Philadelphia, PA: SIAM, 1996.
[7] L. Jähnicke and A. Kost, “Convergence properties of the Newton method
B. TEAM Problem 21 for nonlinear problems,” IEEE Transactions on Magnetics, vol. 34, no.
The discretized geometry of the TEAM Workshop problem 5, pp. 2505–2508, Sept. 1998.
[8] O. Bíró, K. Preis, and K. R. Richter, “Various FEM formulations for
21 is shown in Fig. 3. Since both structures have a symmetry the calculation of transient 3d eddy currents in nonlinear media,” IEEE
plane only a half is computed with about 475.000 unknowns. Transactions on Magnetics, vol. 31, no. 3, pp. 1307–1312, May 1995.

You might also like