Surrogate Modeling Framework for Cost-Effective and Efficient Design of Dielectric Resonator Antennas in Wireless Communication Systems

This paper presents a surrogate-Assisted optimization framework for dielectric resonator antennas (DRAs) operating at 24 GHz, targeting fast impedance matching through automated electromagnetic (EM) simulations and deep neural network (DNN) modeling. Two DRA geometries are investigated: A rectangular DRA (RDRA) and a cylindrical DRA (CDRA). For each geometry, two dielectric constants are considered (\varepsilon_r=10 and \varepsilon_r=79), resulting in four design cases. A CST Studio Suite time-domain EM solver is coupled with Python (CST interface) and MATLAB to generate training datasets over a wide frequency window (18-30 GHz). A DNN surrogate is trained to learn the mapping from key physical design variables (DRA dimensions, substrate size parameters, probe insertion depth, and feed offset location) to the matching objective. Bayesian optimization is then performed on the learned surrogate to rapidly identify candidate designs with minimized |S11| near 24 GHz. The results indicate that high-permittivity DRAs tend to require increased EM simulation time compared to moderate-permittivity DRAs due to shorter effective wavelength inside the resonator and tighter numerical discretization demands. The proposed workflow enables significant reduction in iterative design time and provides a scalable pathway to extend optimization toward multi-objective constraints and larger configurations such as DRA-MIMO systems.