Modelagem neurocomputacional de dispositivos de microondas com estruturas EBG.

Abstract

Lately, there have been intensive studies about the development of periodic structures, such as the photonic crystals, for applications in telecommunication devices. The EBG structures, electromagnetic bandgap structures, also denominated PBG structures, photonic bandgap structures, can be designed on dielectric or metallic materials, through the formation of a grid of holes of which one of the main objectives is to prohibit the propagation of waves in certain frequency ranges. This characteristic has been frequently used in making devices for the optical band as well as for the microwave and millimetric wave bands. The accurate analysis of devices on EBG structures requires the use of numeric methods that demand a substantial computational effort, as the method of the finite differences in time domain (FDTD), among many others. In the last few years, the neural-computational techniques have appeared as powerful and versatile numeric tools for applications in several areas of knowledge. The artificial neural networks, ANNs, show features such as: adaptability, generalization and non-linear, that have contributed to elect them as alternative and advantageous methods for the modelling of several devices of telecommunications. Another characteristic is the significant increase in the processing speed through the use of neural models in simulations of microwave circuits, when compared to physical electromagnetic models. In this work, a new neural computational technique is presented, aiming at reaching an accurate and efficient modelling of EBG structures applied to several devices, such as waveguides, microstrip lines and coupled microstrip lines. This method, denominated Sample Function Modular Artificial Neural Network, SF-ANN modular, makes use of the sample function as an activation function and has a similar configuration to that used in the radial base functions artificial neural networks (RBF-ANN). An excellent agreement is observed between the numerical results obtained in this work and measured values, available in the literature, demonstrating the accuracy of the SF-ANN models. Besides, the SF-ANN models accomplish generalizations for areas of interest, for which there are no available results.