Wideband neural modeling of wire antennas: Feed-forward neural networks versus recurrent ones

In today's communication systems, broadband services play more and more important role. Therefore, communication systems have to be modeled and designed in a wide range of frequencies in order to ensure the proper parameters over the whole band of interest. If numerical models of antennas, transmission lines and circuits are developed in the frequency domain, the structures have to be analysed on each harmonic frequency separately. Exploiting time-domain formulation and narrow-pulse excitation, the information about the structure behaviour can be obtained for wide frequency band
within a single analysis. Since the evolution of an observed quantity is computed iteratively in the time domain, information obtained in previous time samples is exploited for computing successive
values. Therefore, the time-domain analysis shows better effciency.
Although rather effcient, full-wave time-domain numeric models can not be directly incorporated
into the complex design tools because their CPU-time demands are not suffciently low. Therefore, results of the time-domain analysis have to be approximated. Since developing closed-form approximate expressions is time-consuming and requires an intellectual effort, ways of automated approximation were investigated and exploitation of artificial neural networks was proposed. Nevertheless, there are no strict rules recommending the strategy for architecture selection, training method choice, structure
of training patterns in the training set, etc.
In the proposed paper, feed-forward artificial neural networks and recurrent ones are trained in
order to approximate results of the analysis of wire antennas by the time-domain method of moments. Approximation abilities of those two architectures are in detail compared when various training methods and different structure of training patterns is used. On the basis of the comparison results, a methodology for building wide-band neural models of the components of communication systems is worked out. The validity of the methodology is verified when modeling various wire antennas.
In the proposed paper, methodology for automated building of time-domain neural models of electromagnetic structures is proposed. The methodology concentrates on the accuracy and effciency of building models. Various strategies are worked out, which differ in the architecture, training method and the structure of the training set, are in detail compared and their properties are illustrated on testing examples.