Automated electromagnetic antenna synthesis using neural networks and the finite difference time domain

An investigation dealing with the application of neural networks to the synthesis of devices that involve problems which are computational intensive, was completed. The problem involved finding suitable geometries for microwave devices, given a desired electrical performance. In order to perform this task, heteroassociative memories and a randomization procedure of neural network inputs and outputs were utilized. The research presented an original and novel solution to the synthesis process. The subjects used were a printed dipole antenna and a microstrip line. These are widely used by industry in a variety of applications, such as antenna arrays and microwave circuits. A Finite Difference Time Domain (FD-TD) electromagnetic model was developed for the analysis of a printed dipole. This model was verified with measurements and with HFSS, a finite element tool. Practical, yet simple neural networks, were developed for the synthesis of geometrical configurations. The goal of the printed dipole problem was to achieve a VSWR of 2 or less over a 40% bandwidth. The goal of the microstrip line problem was to achieve a line impedance of 50 Ohms. Hence, the microstrip line synthesis represented a trivial problem, used to gain understanding into effective synthesis processes. The input of the microstrip line synthesis neural network was the desired line impedance, and the output was the line width to be determined. The printed dipole synthesis represented a challenging problem, and was used to test the effectiveness of the developed synthesis process, since several input and output parameters were used. The neural network topology uses the delta rule learning algorithm with one hidden layer, and the synthesis process is carefully described in this document. An extensive software development is accomplished in order to achieve the goals set forth herein.