Machine-Learning Surrogate Modeling for $S_{11}$ and Realized-Gain Prediction of a 3.5-GHz 5G Inset-Fed Microstrip Patch Antenna
Year: 2026
Type: conference-paper
Abstract: Repeated full-wave electromagnetic simulations can be time-consuming during early-stage antenna design, especially when many geometrical configurations need to be evaluated. This paper presents a machine-learning surrogate modeling approach for predicting the reflection coefficient$S_{11}$and realized gain of a 3.5-GHz inset-fed microstrip patch antenna for sub-6-GHz 5G applications. A dataset of 100 CST Studio Suite simulations is generated using Latin hypercube sampling over five geometrical parameters: patch length$L_{p}$, patch width$W_{p}$, inset gap, inset depth$y_{0}$, and feed-line width$W_{f}$. The operating frequency is also included as an input feature, enabling the models to learn how the antenna response varies with both geometry and frequency. Six regression models are trained and compared: Random Forest, Extra Trees, LightGBM, XGBoost, CatBoost, and multilayer perceptron (MLP). The MLP achieves the best prediction accuracy, with$\mathbf{R M S E}=0.3463 \mathbf{d B}$and$R^{2}=0.9927$for$S_{11}$, and$\text{RMSE} =0.1899 \mathbf{d B}$and$R^{2}=0.9963$for realized gain. The proposed approach provides a fast and practical tool for estimating antenna performance during early-stage design exploration.
Institution: Abdelmalek Essaâdi University
Language: English
Topic: Antenna Design and Analysis
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204140649