Wide band THz MIMO antenna with isolation prediction using machine learning algorithms for IoT applications

A compact wideband two-port MIMO antenna has been developed to function in the Terahertz (THz) frequency band, specifically at 1.58 THz and 2.6 THz, with the aim of enhancing Internet of Things (IoT) applications. This antenna is constructed using a polyimide substrate with graphene patch radiator and copper ground plane. The antenna design progresses through four stages: rectangular patch, corner-chamfered octagonal patch, ring-frame with central aperture, and finally, cross-slot (plus-sign) loaded octagonal structure. The final design demonstrates two resonances at 1.6 THz (S11 = -32 dB) and 2.35 THz (S11 = -53 dB).

To optimize the MIMO configuration, three placement methods were considered, resulting in an inter-port isolation of -18 dB in the 2.6 THz range and -25 dB to -45 dB in the 1-1.5 THz range. The antenna also exhibits a gain of 6-6.5 dBi across the operating bands. The diversity gain is reported as 9.9-10 dB, while the channel capacity loss is less than 0.5 bits/s/Hz, and the total active reflection coefficient is less than -10 dB.

To address the computational burden of extensive parametric sweeps, a supervised machine learning (ML) surrogate model was created to predict inter-port isolation (S21) as a function of frequency and inter-element spacing. This ML approach significantly reduces the computational load of repeated full-wave HFSS analysis. The gradient boosting regressor, after hyperparameter optimization using Optuna, achieved an R2 of 99.99%, mean absolute error (MAE) of 0.26 dB, and root mean square error (RMSE) of 0.37 dB for the test data.

This model outperformed all baseline regression methods, making it a reliable surrogate for THz MIMO antenna isolation prediction.