Image Classifier Based Generative Method for Planar Antenna Design

TL;DR

The paper addresses rapid PCB antenna prototyping for wearables by decoupling geometry into fixed-dimension rectangles and using random-position statistics to evaluate dimension sets. It introduces a CNN-based image classifier as the core of a generative method, rendering geometry as a 2D image with coordinate channels and predicting a design-score and an EM-response proxy such as $S_{11}$, while employing surrogate scoring and regularization to minimize EM simulations. The dimension selection uses median scores over many random placements to identify promising shapes, followed by a classifier-guided placement loop with adaptive thresholds and optional trust-region optimization to refine prototypes. On two wearable examples, including a compact AR-glasses antenna, the workflow yields realistic prototypes whose measured performance aligns with simulations and is not inferior to designs from experienced engineers, demonstrating a practical, AI-assisted approach for rapid, expert-free antenna prototyping.

Abstract

To extend the antenna design on printed circuit boards (PCBs) for more engineers of interest, we propose a simple method that models PCB antennas with a few basic components. By taking two separate steps to decide their geometric dimensions and positions, antenna prototypes can be facilitated with no experience required. Random sampling statistics relate to the quality of dimensions are used in selecting among dimension candidates. A novel image-based classifier using a convolutional neural network (CNN) is introduced to further determine the positions of these fixed-dimension components. Two examples from wearable products have been chosen to examine the entire workflow. Their final designs are realistic and their performance metrics are not inferior to the ones designed by experienced engineers.

Figures (18)

• Figure 1: Overall workflow to design PCB antennas using basic components.
• Figure 2: A CNN-driven image classifier functions as both a predictive score discriminator and a part of the PCB antenna prototype generator.
• Figure 3: Structure of the CNN model. Source code can be find in the repository https://github.com/yz564/antenna_ai.
• Figure 4: Workflow of the dimension selector.
• Figure 5: Comparison of the sample score distribution for five sets of dimensions. (a) Each '+' sign represents one simulation data. (b) Each curve represents 100 random position simulation data with the fixed dimensions.