Design optimization of a C/X-band metasurface–DGS microstrip antenna using machine learning

Microstrip patch antennas (MPAs) typically suffer from limited bandwidth and low gain in C- and X-band applications. This study presents a machine learning (ML)-assisted antenna design that integrates an elliptical metasurface with a defected ground structure (DGS) to enhance performance. The antenna is implemented on a low-loss F4B substrate (εᵣ = 2.65, tan δ = 0.0015, thickness = 1.6 mm) and excited using a 50 Ω coaxial probe feed. A data-driven optimization framework is developed using two supervised regression models: an artificial neural network (ANN) and a random forest (RF) regressor. These models are trained on 100 full-wave electromagnetic simulations (CST Microwave Studio) to predict return loss (S 11 ), impedance bandwidth, voltage standing wave ratio (VSWR), gain, and directivity. Both models achieve a coefficient of determination (R 2 ) greater than 0.97 and a mean absolute error below 1.5%, demonstrating high predictive accuracy. The integration of ML with electromagnetic design is a key strength of this work. The ML-optimized antenna exhibits simulated performance of −53 dB return loss, 47.6% impedance bandwidth (5.68–9.24 GHz), VSWR of 1.00, peak gain of 4.91 dBi, and radiation efficiency of 98.8%. Compared with similar metasurface-based MPAs, this represents a bandwidth improvement of 17–40%. The ML-assisted approach reduces optimization time by approximately 88% relative to conventional parametric sweeps. All results are simulation-based; experimental validation has not yet been performed. The proposed design is suitable for broadband applications including 5G communication, satellite links, and radar systems. Future work will focus on experimental validation, multilayer gain enhancement, and extension to millimeter-wave frequencies.