Machine Learning-Assisted Quasi-Bisection Method for Pixelated Patch Antenna Bandwidth Optimization

A machine learning-assisted quasi-bisection method (MLAQBM) is proposed for the broadband optimization of the pixelated patch antennas. Distinct from traditional MLA optimization methods, MLAQBM employs a coarse-to-fine pixel partition and machine learning optimization to iteratively refine the shape of the antenna. Two implementation strategies are introduced. In the first strategy, larger pixels are initially partitioned and samples are collected. A surrogate model is then established using convolutional neural networks (CNN), and an initial shape is derived through particle swarm optimization (PSO). Again, further refinement is achieved by finer pixel partitioning and subsequent CNN-based PSO. The second strategy leverages a pre-existing broadband E-shaped patch design, which is then subdivided into smaller pixels. By collecting samples, training a CNN surrogate model, and performing PSO optimization, a broader bandwidth is realized. Compared with the reported MLA optimization methods, MLAQBM establishes a stable surrogate model that is robust, easy to converge, and adaptable to different optimization objectives. To validate the proposed strategies, the optimized pixelated patch antennas are fabricated and measured. The measured results demonstrate that the bandwidth of the proposed patch antennas achieves 24.5% and 32.5%.