Material-aware inverse design of rectangular dielectric resonator antennas using machine learning

Abstract

This paper presents a fully automated, data-driven pipeline for the joint dielectric-geometric inverse design of rectangular dielectric resonator antennas (RDRAs) operating at 13.5 GHz in the super high frequency band. Surrogate-based design approaches conventionally adopt artificial neural networks; here, three supervised regression families, multilayer perceptron, random forest and gradient boosting, are instead benchmarked under a common 10-fold cross-validation protocol, using the mean absolute error of the predicted resonant frequency as the selection criterion. The selected model, achieving R2=0.999 for resonant frequency, is extended to a multi-output architecture that predicts five antenna performance metrics, resonant frequency ( fres), minimum return loss ( S11,min), impedance bandwidth, complex input impedance (Re( Z11), Im( Z11)) and voltage standing-wave ratio, from four co-design variables: the relative permittivity ( εr) of a composite material resonator and its three geometric dimensions ( a, b, d). A differential evolution optimizer drives the inverse-design stage, returning diverse, multi-objective-satisfying geometries that were independently validated against full-wave ANSYS HFSS simulations, achieving a mean resonant-frequency error below 10 MHz. Treating permittivity as a continuous material-design variable, rather than as a fixed parameter, the framework jointly explores candidate material–geometry combinations and directly links electromagnetic specifications to the corresponding material-fabrication process. Beyond the RDRA case study, this benchmark-first, material-aware methodology is presented as a reproducible template for surrogate-assisted inverse design of other parametrized electromagnetic structures.