Accurate Modeling by Convolutional Neural-Network Regression of Resonant Frequencies of Dual-Band Pixelated Microstrip Antenna

The modeling of the performance characteristics of pixelated microstrip antennas poses special difficulties as their geometries cannot be readily parametrized compared with the geometries of antennas with standard shapes. A methodology is presented for accurately modeling the two resonant frequencies of a dual-band pixelated microstrip antenna based on convolutional-neural-network (CNN) regression that takes a representation of the entire pixelated surface of the antenna as input. The predictive performance of the CNN and shallow and deep conventional feedforward neural-net architectures were compared; architectural and learning algorithm hyperparameters were determined by means of Bayesian optimization. The CNN achieved the best predictive results of all the networks with mean relative errors on test predictions of 0.13% and 0.22% for the two resonant frequencies, respectively.