Data-Driven Classification of Reflection Coefficients for Pixelated Microstrip Antenna Design

Abstract

This paper presents a machine learning-based framework to classify the reflection coefficient $\left(\left|S_{11}\right|\right)$ of pixelated microstrip patch antennas as either below or above -10 dB. The antenna geometry is modeled as a $7 \times 5$ binary matrix of removable pixels, enabling a highly flexible design space. A dataset of 7,995 antenna configurations was generated through full-wave FEM simulations in Ansys HFSS and automated using the PyAEDT Python interface. A Multi-Layer Perceptron (MLP) was trained across the nine evaluated frequencies 1-9 GHz and achieved f1-scores exceeding 82% for all cases. The proposed method significantly reduces the computational cost associated with traditional full-wave optimization and provides a fast, datadriven tool for the design of pixelated microstrip antennas.