Generative Deep-Learning Design of Single- and Dual-Polarized Four-Port MIMO Antennas for Full-Duplex Communication

This work presents a machine-learning-driven framework for designing single- and dual-polarized, four-port MIMO antennas for in-band full-duplex (FD MIMO) operation. A symmetry-constrained, pixel-based layout enables efficient exploration of the design space. A multi-task convolutional neural network (CNN) surrogate is trained to jointly predict return loss and multi-port coupling from symmetric pixel layouts, accurately modeling the amplitude of the$4\times 4$S-parameter matrix across 5–8 GHz. By embedding this surrogate within a binary genetic algorithm, the framework synthesizes antenna layouts that simultaneously achieve impedance matching and high inter-port isolation across the operating band. Fabricated prototypes show close agreement with predictions and simulations, achieving a return loss greater than 10 dB, isolation up to 36.25 dB, and realized gain of 7.7 dBi at 6.2 GHz. Unlike prior multi-port designs, the approach delivers low loss and robust isolation without relying on additional decoupling structures such as defected ground planes, resonators, or stacked patches, establishing a scalable methodology for next-generation FD MIMO antenna systems.