Machine Learning based Surrogate Modeling for Wave Impedances in Rectangular Dielectric Waveguides

Even though the behavior of rectangular dielectric waveguides (RDWGs) can be evaluated accurately using numerical simulations, limited time and computational resources generally do not allow an extensive exploration of the design space. Thus, in this paper we present a way to develop a fast-to-evaluate surrogate model for the wave impedance of the fundamental mode HE11 in RDWGs at frequencies between 134 and 141 GHz. To this end, gaussian process regression (GPR) with a customized covariance function as well as artificial neural networks (ANNs) and polynomial regression (PR) are applied. Compared by the mean absolute deviation (MAD), our finally proposed model shows much better agreement with numerical solutions than Maractili’s famous approximate analytical approach (MA). In contrast to MA, our approach is furthermore applicable to DWGs with arbitrary cross-sections.