Machine Learning-Driven Synthesis of Anisotropic Antenna Arrays for Directional Beamforming Optimization

To address the limited design freedom caused by the assumption of isotropic elements in traditional antenna array design, a machine learning-assisted array synthesis method is proposed. By constructing a surrogate model for anisotropic active elements, the method combines Gaussian process regression to predict the active element patterns under arbitrary spatial arrangements, geometric parameters, and complex electromagnetic environments, effectively expanding the array design space. Using a 12-element linearly coupled patch antenna array as a validation example, under constraints of uniform amplitude and 1-bit discrete phase quantization excitation, genetic algorith-mbased joint optimization of element size and rotation achieves radiation characteristics with a target beam direction deviated by 50° from the normal. Experimental results demonstrate that the method suppresses symmetric beams while reducing the side lobe level below -10 dB, and the prediction results show high consistency with full-wave simulations, validating the feasibility and accuracy of machine learning-assisted anisotropic element array design.