Machine learning-based design and optimization of a cross-shaped circularly polarized multiband patch MIMO antenna for 6G sub-THz communication systems

Abstract

This paper presents a machine learning-assisted design and optimization of a multiband multiple-input multiple-output (MIMO) patch antenna for sub-terahertz (sub-THz) communication systems. The antenna is optimized using Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Differential Evolution (DE). A set of machine learning models is evaluated, and the best-performing model is selected based on prediction accuracy, convergence time, mean absolute error (MAE), and coefficient of determination ( \(R^{2}\) ), and is subsequently employed for surrogate-based optimization. To enhance gain and directivity, multiple array configurations, including 1 \(\times\) 4 linear, reverse, and 2 \(\times\) 2 planar MIMO structures, are investigated. The proposed antenna employs an asymmetric cross-shaped slotted patch to achieve circular polarization and is designed on a Rogers RT5880 dielectric substrate. The antenna operates at 134 GHz, 230 GHz, and 450 GHz, achieving bandwidths of 5.98 GHz, 14.64 GHz, and 82 GHz, respectively. The single-element antenna provides a gain of 6.78 dBi, while the linear array achieves a peak gain of 13.9 dBi. The proposed design also attains radiation efficiency of up to 80% and high isolation below –30 dB at operating frequencies without using a defected ground structure. Circular polarization is confirmed by an axial ratio below 3 dB over the operating bands. Among the evaluated machine learning models, Random Forest achieves the best performance with an \(R^{2}\) value of 98.18%, demonstrating the effectiveness of the proposed surrogate model. The machine learning-assisted surrogate optimization significantly reduces computational time, computational complexity, and resource requirements while maintaining high prediction accuracy. Unlike existing studies that primarily focus on antenna performance prediction, this work proposes a machine learning-assisted multi-objective optimization framework for the design of multiband circularly polarized MIMO antennas operating in the sub-THz regime.