ML assisted optimization of DR based MIMO antenna for 5G communication systems

Abstract

Abstract This work presents a dual-port cylindrical dielectric resonator antenna (CDRA) based MIMO design for 27-GHz 5G communication, assisted by a suspended complementary split-ring resonator (CSRR) metasurface for port decoupling. The antenna is realised on Rogers RT Duroid 5880 substrate of thickness 0.254 mm, and the cylindrical dielectric resonator uses Rogers RT Duroid 6010 material. In simulation, the metasurface improves the mutual coupling from about − 27.56 dB to − 45.22 dB at 27 GHz (≈ 17.6 dB enhancement) while maintaining impedance matching. A prototype is fabricated, and measurements confirm an operating band of 26.24–27.94 GHz and show clear isolation improvement when the metasurface is applied, with a peak realised gain of around 5 dBi in the band. To reduce optimization time at millimetre-wave frequencies, a dataset generated in HFSS (19,100 samples) is used to train surrogate regression models, including Decision Tree, K-Nearest Neighbours, Random Forest, Extreme Gradient Boosting, and a Deep Neural Network (DNN), to predict S 11 and S 12 from geometry parameters and frequency. Among the studied models, the DNN gives the best prediction for S 11 (R 2 = 0.991), while Random Forest and DNN show the best performance for S 12 (R 2 = 0.993 and 0.990). The combined antenna design and surrogate modelling approach supports faster parameter search for high-isolation mmWave DRA-MIMO antennas.