Neural network applications in microwave device design

Abstract

This paper deals with many different neural network architectures that have been introduced to simulate the electromagnetic behavior of complex microwave devices and antennas. Many issues linked to the peculiarities of such devices are addressed. Furthermore, the inverse problem of designing a microwave device, once the required characteristics are given, is also developed by using a neural network approach.

A carbon-nanotube-based frequency-selective absorber

“…A finer discretization would have led to a more complete look-up table but would have been too time consuming, and hence an interpolation of these values has been performed through an artificial neural network (ANN) trained to this aim and outputting the expected admittance of the FSS as a function of the variable geometrical and electrical parameters. ANNs are known to be exceptional approximators [18] and have been successfully used in electromagnetism in many different applications [19, 20].…”

“…ANNs are known to be exceptional approximators [18] and have been successfully used in electromagnetism in many different applications [19,20]. In particular, analyses have been done via full-wave FEM for 1 1 , 1 2 [ [1,1.5,1.7,2] and ring radii r 0 + 15% The analysis has been carried out for a discrete frequency set (6-16 GHz with 1 GHz step) and a discrete set of angle of incidence (0-608 with 58 step).…”

Recent advances in neural network‐based inverse modeling techniques for microwave applications

“…For inverse modeling with the challenges due to the nonuniqueness problem, if the problem is in the low-dimensional spaces, both the comprehensive neural network inverse modeling methodology 20 and the multivalued neural network inverse modeling technique 41 are applicable; if the problem is in the medium-dimensional spaces, the multivalued neural network inverse modeling technique 41 becomes a suitable choice. For the inverse problems without nonuniqueness problem, if the inverse model has low dimensional inputs so that the input-output relationship is able to be represented by shallow ANNs effectively, the shallow ANN can be directly trained to build the inverse model 34 ; if the inverse model has high-dimensional inputs which exceed the capabilities of the shallow ANNs, the hybrid deep neural network technique 42 is a good choice. For the inverse problems with both nonuniqueness problem and high-dimensional inputs, one possible way is to combine the hybrid deep neural network technique with the multivalued neural network inverse modeling technique, which can be a very interesting future direction in neural network-based inverse modeling.…”

“…Researches in inverse modeling using ANNs have been reported. An ANN inverse model is trained in the direct inverse modeling technique, 34,35 where the inputs and outputs of the forward problem are swapped to be the training data for the inverse model. The neural inverse space mapping technique, which makes use of the inverse of the mapping from the fine to the coarse modeling parameter spaces, 36,37 has been used in EM-based optimization and design.…”

Artificial Neural Networks for Wedge Diffraction Coefficients

“…Indeed, ANNs are a useful software tool representing a "universal approximator", that is a black box able to reproduce the behavior of any given system with an arbitrary accuracy [3]. Among the most interesting features of the ANNs there is that they can leam from known examples, namely they can be automatically trained to fit a given function on the basis of a set of inputs and corresponding desired outputs the ANN should match [4,5]. The ANN training time is usually long, since several exact diffraction coefficients for many configurations must be computed, possibly involving different approaches [1,2].…”