Optimizing microstrip patch antenna performance: Leveraging machine learning for comparative analysis of antenna di-mensions and slot structures

• This paper presents a focused and comprehensive literature overview of conventional computational antenna designs and optimizations. In complex antenna structures, manual exploration may be impractical and time-consuming. In the case of large datasets, this capability becomes particularly valuable when using machine learning algorithms. The insights into antenna physical design are based on electromagnetics and numerical methods, with a focus on their implementation in machine learning algorithms. • Utilizing a simple example of Classification Machine Learning in MATLAB for antenna design represents a paradigm shift from traditional methodologies, particularly in the context of rectangular antenna designs. The incorporation of classification models is adept at discerning intricate patterns within datasets, enabling accurate predictions of resonance frequencies and the subsequent generation of antenna designs with enhanced performance metrics. The achieved accuracy of approximately 88.3% underscores the efficacy of these advanced techniques in optimizing antenna parameters and predicting resonance frequencies with a high degree of precision. Microstrip patch antennas (MPAs) are widely used in wireless systems due to their compact size and ease of fabrication, but their design still relies on time-consuming full-wave electromagnetic simulations. This study presents a unified machine learning (ML)-assisted pipeline for predicting and optimizing antenna performance across three common MPA shapes—rectangular, circular, and triangular—with various slot configurations and substrate types. Using a curated dataset of 1,500 designs, we apply support vector regression (SVR), which models nonlinear relationships using kernel functions; Gaussian process regression (GPR), which provides both predictions and uncertainty estimates; and artificial neural networks (ANN), which learn complex patterns from data. For classification tasks, we introduce a lightweight convolutional neural network (CNN) to pre-screen slot regimes. We also propose a smart simulation strategy that uses model uncertainty to trigger new simulations only when needed, reducing computational cost. Validation includes parity plots, residual histograms, and S11 overlays, while sensitivity analysis highlights the most influential geometric and material features. The pipeline offers interpretable, simulation-efficient design guidance and lays the foundation for future hardware validation and multi-objective optimization.