A Two-Stage Machine Learning Framework for Predicting and Validating the Structural and Beam Characteristics of a Wideband Switched-Beam Antenna

Abstract

This paper presents a two-phase machine-learning (ML)-based approach for the design and performance forecasting of a wideband circular microstrip switched-beam antenna operating at the 2.6 GHz mid-band for 5G communication. In Phase 1, five supervised learning algorithms Decision Tree (DT), Random Forest (RF), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Support Vector Regression (SVR) are employed to model the relationship between the antenna’s geometric parameters (inner radius, outer radius, and slot gaps) and its resonant characteristics. The DT model achieves the highest prediction accuracy when benchmarked against CST full-wave simulation results, enabling reliable estimation of center frequency and impedance bandwidth without repeated electromagnetic simulations. In Phase 2, three machine learning models DT, RF, and KNN are trained using CST-generated datasets to predict the main-beam direction and bandwidth of the switched-beam structure based on the number, location, and distribution of short-circuited holes. The DT model again outperforms the others, achieving 94.56% accuracy in beam-direction prediction and 99.03% accuracy in bandwidth prediction. A prototype antenna fabricated on an FR-4 substrate demonstrates four-direction beam switching (0°, 90°, 180°, and 270°) with a measured impedance bandwidth of 380 MHz (2.48–2.86 GHz), closely matching both simulation and ML predictions. Compared with conventional trial-and-error simulation-based design workflows, the proposed framework significantly reduces antenna design time while maintaining high prediction accuracy, demonstrating the practical effectiveness of ML-assisted design for compact, cost-effective, and energy-efficient 5G antennas.