Design and Optimization of Differentially-Fed Microstrip Cross-Grid Array Antenna With DNN Surrogate Model Enabled AGA

In this paper, we firstly present a kind of microstrip cross-grid array antenna (MCGAA) with a differentially-fed excitation for millimeter wave (mm-Wave) applications. The incorporation of a novel cross-grid element enables its more compact array arrangement, thereby achieving high aperture efficiency (AE). Besides, we further introduce a set of parasitic patches between two adjacent radiated elements to broaden the operating bandwidth (BW) of antenna, which is demonstrated by both simulation and experiment. Further, based on such an array structure, one deep learning-enabled optimization scheme is developed instead of the traditional full-wave electromagnetic (EM) based optimization procedure, realizing fast design of a larger array scale with low time cost. In particular, one deep neural network (DNN) is employed as a forward solver, which is trained by using 4000 datasets with labels. The final training accuracy is about 97.10%, while the testing accuracy of in-range and out-of-range data are 94.35% and 90.62%, respectively. Furthermore, an adaptive genetic algorithm (AGA) is exploited to accomplish target-to-structure design process. For verification, we design and optimize a larger antenna array than either of the training samples using the proposed design method. Both the simulated and measured verification results are presented with the desired design targets satisfied, demonstrating both effectiveness and scalability of the proposed method. Owing to the performance merits and efficient optimization strategy, it is highly expected that the proposed array structure and design method will be suitable and useful for the fifth-generation (5G) mm-Wave applications.