Deep Learning Aided Modelling and Inverse Design for Multi-Port Antennas

With the prevalence of multiple-input multiple-output (MIMO) systems, multi-port antenna design has become an important research area. In this work, we approach the multi-port antenna design problem to accelerate the design cycle, expanding design space, and finding non-intuitive designs that can potentially yield better performance than existing template-based designs. To achieve these, we rely on the optimization of a discretized surface, which can implement near-arbitrary antenna shapes. However, performing an electromagnetic (EM) optimization with a large number of variables is prohibitively costly. On the other hand, if EM simulations can be replaced by a machine learning (ML) based approach, antenna optimization could be accelerated greatly. To this end we utilize a convolutional neural network (CNN) for the modeling of multi-port pixelated structures. A genetic algorithm (GA) in conjunction with CNN is used to perform inverse design. Example designs for various optimization targets have been shown in support of the proposed approach.