Transfer Learning Based Rapid Design of Frequency and Dielectric Agile Antennas

Abstract

Deep learning frameworks are gaining prominence in the electromagnetics community for designing microwave and mm-wave devices. This paper presents a computationally efficient transfer learning technique for designing and scaling multi-band microstrip antennas to a desired dielectric and frequency of interest. The proposed methodology involves a two-step process. First, a pre-trained model trained extensively on air-filled microstrip antennas is used for knowledge transfer. This pre-trained model is fine-tuned with a limited set of dielectric simulations, reducing data acquisition costs. In the second step, the developed forward model serves as a surrogate to design dielectric-filled antennas using the Improved Binary Particle Swarm Optimization algorithm. In contrast to conventional methods, this approach enables the design of compact antennas across various dielectrics and frequency ranges, with a significantly reduced number of time-consuming dielectric simulations (88% fewer simulations) and a lower neural network training time (75% lesser time). We analyze the optimal ways of generating dielectric antenna datasets via scaling, and perform sensitivity analysis with respect to the antenna's physical parameters. We report simulation and experimental results for single and double band antennas fabricated using the proposed approach.