Deep Learning-Inspired Linear Regression Technique for Accurate Microstrip Antenna Performance Analysis

This paper presents a computationally efficient Linear Regression (LR) model for predicting the performance of microstrip patch antennas using key geometrical design parameters. While deep learning models have recently gained traction in antenna modeling due to their ability to learn complex nonlinear relationships, they often require large datasets, extensive training time, and high computational resources. In contrast, the proposed LR-based approach offers a lightweight and interpretable solution for performance prediction in early-stage antenna design. The model predicts two critical antenna metrics Gain and S11 (return loss) from six geometrical parameters: lu2, scale1, scale2, scale3, ud1, and uy1. The dataset used for training was derived from HFSS-based full-wave simulations and validated through VNA testing of a fabricated Y-shaped microstrip antenna operating in the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3.25-4.25 \text{GHz}$</tex> range. The LR model demonstrated high prediction accuracy, with a Test MAE of 0.0032 and MSE of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$5.206 \times 10-5$</tex> for Gain, and MAE of 0.01 and MSE of 0.0004 for S11. The <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3.25-4.25 \text{GHz}$</tex> band is widely used in sub-6 GHz 5G networks, offering an optimal balance between coverage and data throughput for high-speed wireless communication. These results highlight that LR can serve as a scalable and low-cost alternative to deep learning models, particularly in scenarios where real-time performance estimation and rapid design iterations are required.