A Novel Deep Learning Framework With Separate Networks for Robust Antenna Array Synthesis

The synthesis of antenna array excitations from a desired radiation pattern presents a challenging inverse problem in electromagnetics. While deep learning offers a promising solution, existing approaches often employ a single network for the joint prediction of both amplitude and phase, which can be suboptimal. This paper introduces a novel dual‐network deep learning framework that employs two separate and specialized networks to independently and simultaneously learn the mappings for amplitude and phase. This architecture allows a simpler network to be dedicated to the less complex amplitude profile, leading to faster convergence and a reduced computational load. To enhance robustness against heavy‐tailed noise and prevent overfitting, the correntropy loss function is adopted instead of conventional MAE or MSE. Furthermore, a cosine annealing learning rate schedule is implemented to achieve superior and faster convergence compared to fixed, stepwise, and exponential decay schedules. The proposed model is rigorously evaluated through extensive Monte Carlo simulations (100 runs) to ensure statistical reliability. Simulation results for a linear array demonstrate that the proposed method provides a favorable trade‐off between reconstruction accuracy and computational efficiency compared with conventional techniques. The separate network structure alone reduces training and test error by approximately 11%, while the correntropy loss yields a further 32% reduction in the radiation pattern reconstruction (RMSE) compared to MAE loss. This framework paves the way for highly accurate and real‐time reconfigurable antenna systems.