Machine Learning-Assisted Synthesis of Heterogeneous Arrays via Active Geometric Design

Mutual coupling often leads to discrepancies between intended and synthesized array patterns due to the distortion of active element patterns (AEPs). In array synthesis, existing machine learning methods are predominantly confined to predicting AEPs for fixed structures, imposing a passive constraint that limits optimization flexibility. A holistic framework is introduced for the active geometric design of heterogeneous arrays, in which the geometric parameters of individual elements are treated as degrees of freedom alongside traditional ones. To manage the complexity associated with mutual coupling, a variational autoencoder-based surrogate model is developed to efficiently capture the mapping between diverse element structures and their radiation patterns in a latent space. Numerical comparisons indicate a 9.8× speedup in prediction over traditional methods. Furthermore, under rigorous uniform amplitude and 1-bit phase constraints, sidelobe suppression improvements of over 8 dB are achieved at scanning angles compared to homogeneous baselines, validating the efficacy of incorporating element-level geometric design into the global optimization process.