T h i s paper discusses the application of Neural Network ( N N ) technique in the m,odelling of a typical electromagnetic (EM) field problem, consisting of a coaxially-driven dielectric-ring resonator antenna. T h e N N takes in consideration the geome2:rical and material parameters of the EM m.odel, and the reflection coeficient values f o r different frequencies in the bandwidth of interest. T h e N N i s trained off line using training sets generated by the j n i t e element method. Once a n acceptable N N model i s found by :iimulation it will finally be implemented in hardware in order t o provide the rea.1 t i m e behaviour requested f o r further integration in a n EM-CAD tool. 1: IntroductionIn the solution of electromagnetic (EM) field problems various numerical methods have been in use. Numerical methods widely used include the finite element method (FEM), the finite difference time domain (FDTD) method, and the Method of Moments (MOM).In typical EM field problems the solution by numerical techniques generally involves three stages. The first stage is the problem formulation which mainly involves the writing up of the system equations to be solved. The second stage is the creation of the geometry that describes the problem and the necessarry discretization of the problem domain into smaller elements. The third stage is to solve the system of equations using the method of choice.There are various situations where repeatitive computation of EM field is required, such as in optimization of the problem geometry for optimum outputs. As a matter of fact, such a need arises in several CAD tools for very high speed circuit design. A minor change in the problem geometry would require a different discretization which is itself a time consuming exercise. In addition, the solution of the system of equations takes a lot of computer time depending on the complexity of the problem and the resulting number of equations.The existing CAD packages such as SPICE and Libra would need to integrate EM field modelling tools, but that cannot be done in real-time owing to the inherent limitation of the known numerical methods as described cl r b ove.It is the purpose of this paper to investigate the feasibility of using Neural Network (NN) as a parallel modelling device that would help integrate E M field models with real-time CAD tools.As an example, a discussion of the implementation of NN to a problem solved by FEM will be provided. The problem to be solved, is an axisymmetrical problem 0-8186-7456-3/96 $5.00 0 1996 IEEE 387

No abstract