Impact analysis and optimization of notched high-gain Antipodal Vivaldi antenna using machine learning

Abstract

Optimizing antenna parameters without knowing their impact on the gain is a time-consuming and iterative process. This article shows how the Classification and Regression Tree (CART) technique, the Random Forest (RF) algorithm, and the Principal Component Analysis (PCA) method is used to study the effect of geometrical parameters of rectangular notches of the Antipodal Vivaldi antenna on gain. Carving out the notches is a general approach for end-fire antenna design. Integrating the EM simulation with machine learning could reduce the time spent on AVA design. In this approach, the gain category of the antenna under design: Low, Medium, and High, is first classified by the three multiclass machine learning models, K-Nearest Neighbor (KNN), Artificial Neural Network (ANN), and Support Vector Machine (SVM). When the predicted category of gain belongs to the High class, in that case, the four regression models, Polynomial, SVM, ANN, and RF, predict the gain value to anticipate the antenna gain according to the notch dimensions to reduce the optimization time of the simulation-driven antenna. In this work, the band-limited approach of notches on gain is also been validated using simulation before fabrication. Machine learning models are optimized the antenna parameters to achieve a high gain within the frequency range of 20 to 40 GHz. A peak simulated gain of 19 dB is achieved at the frequency of 27 GHz. The proposed structure is fabricated utilizing a single-layer Rogers substrate (RO4003C) with a dielectric constant of 3.38 and compact dimensions of 96 mm × 174 mm × 0.508 mm, i.e., 0.32 × 0.58 × 0.0017 λ03; here the symbol λ0 denotes the wavelength of free space at the minimum operating frequency.