Abstract
This work presents a learning-based surrogate-assisted approach for the synthesis and optimization of irregular linear antenna arrays. The proposed methodology addresses grating-lobe suppression in arrays with inter-element spacings exceeding one wavelength, accounting for mutual coupling while simultaneously controlling side-lobe levels (SLL), beamwidth, and spacing constraints. Full-wave simulations of ten-element dipole arrays operating at 26 GHz were performed in High Frequency Simulation Software (HFSS) for random inter-element spacings ranging from 0.5 to 2.5 wavelengths, generating approximately 13,000 array configurations. A multilayer perceptron (MLP) was trained to map the normalized inter-element spacings to the complex radiated electric field, providing a fast surrogate model for radiation-pattern prediction. The trained model achieved a validation mean squared error (MSE) of approximately 17×10 −3 on the normalized complex electric field radiated components, demonstrating excellent agreement with HFSS reference patterns, quantified by a mean absolute error (MAE) of 0.1 V, calculated on the test dataset. The surrogate model was then embedded into a Differential Evolution optimizer to perform a 9-dimensional search for optimal inter-element spacings in array geometries satisfying prescribed radiation-pattern masks. The optimization results demonstrate effective suppression of side and grating lobes under different spacing constraints and SLL targets.