Deep learning-based S-parameter prediction of a progressively extended eight-port terahertz MIMO antenna for 6G and next-generation IoT

Abstract

Terahertz (THz) antenna design for sixth-generation (6G) wireless systems is computationally intensive owing to the strong sensitivity of radiation performance to geometrical and material parameters. This paper proposes a deep-learning-assisted framework for the progressive design of a compact multiband THz antenna, evolving from a single-port radiator to an eight-port multiple-input multiple-output (MIMO) configuration. A deep neural network (DNN) is trained on 130,000 parametric samples to predict the real and imaginary parts of the reflection coefficient, replacing repetitive full-wave simulations. Stochastic Gradient Descent (SGD), RMSProp, and Adam optimizers are comparatively evaluated, with Adam yielding the best performance, achieving a mean squared error (MSE) of 0.003015 and a coefficient of determination ($R^2$) of 0.962736. The DNN-optimized antenna exhibits stable multiband resonance over the 0.3–0.8 THz frequency range. When progressively extended to an eight-port MIMO array, the design achieves a return loss better than −22.5 dB, low inter-element mutual coupling, an envelope correlation coefficient (ECC) below 0.0017, and a diversity gain (DG) of 10 dB. The proposed DNN-driven framework substantially reduces the need for repeated full-wave simulations and provides an efficient methodology for the design of compact THz MIMO antennas for future 6G wireless applications.