A Low-Cost Optimization Method for 2-D Antennas Using a Disassemblable Convolutional Neural Network

To enhance antenna optimization efficiency, a common method involves constructing a surrogate model for the antenna. This model establishes a rapid mapping relationship between the antenna’s design variables and its performance. Such a surrogate model can replace full-wave simulations, thereby obtaining low-cost optimization. This article proposes a surrogate model based on machine vision technology to achieve efficient antenna optimization. Unlike traditional models, this surrogate model does not accept design variables of the antenna as input but uses 2-D images depicting the antenna instead. Different pixel values in the images represent various antenna materials, and regions formed by identical pixel values denote the antenna’s geometric dimensions. Leveraging the powerful image processing capabilities of convolutional neural networks (CNNs), this surrogate model can learn the relationship between the geometric parameters and material properties of the antenna and its performance. In addition, to further enhance the evaluation accuracy of the surrogate model, modifications are made to the traditional CNN, allowing the surrogate model to be decomposed into multiple parts for use. The experimental results indicate that this machine vision-based disassemblable surrogate model can significantly improve the evaluation accuracy of antenna design solutions and reduce the dependence on full-wave simulation for optimization.