Machine-Learning Surrogate-Assisted Differential Evolution for S11 and Bandwidth Optimization of a Reconfigurable Microstrip Patch Antenna
Year: 2026
Type: conference-paper
Abstract: Full-wave electromagnetic optimization of reconfigurable microstrip antennas is computationally expensive, since each candidate geometry requires an accurate numerical solver evaluation. This paper proposes a machine-learning surrogateassisted differential evolution framework for efficient optimization of a sub-6 GHz reconfigurable microstrip patch antenna for wireless communication applications. The antenna geometry is characterized by four parameters:$L, L_{0}, W$, and$W_{0}$. A CST Studio Suite dataset of 75 antenna geometries is generated over the 4-7 GHz frequency range. Six regression models are evaluated as surrogates for predicting the full$S_{11}$response, namely linear regression, k-nearest neighbors, random forest, gradient boosting, LightGBM, and multilayer perceptron (MLP). The MLP achieves the best prediction accuracy, with a validation RMSE of 0.4078 dB and an$R^{2}$of 0.9958. The selected surrogate is integrated with differential evolution to maximize the contiguous impedance bandwidth under a -12 dB threshold. The optimized design is validated using an independent CST simulation, yielding a surrogate prediction error of 0.32 dB MAE and 0.51 dB RMSE. The CST validation confirms that the optimized antenna achieves an impedance bandwidth of approximately 1 GHz at the -10 dB threshold. The proposed framework significantly reduces the need for repeated full-wave simulations while maintaining high prediction accuracy.
Institution: Abdelmalek Essaâdi University
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204141602