Low-Sidelobe Millimeter-Wave Series-Fed Patch Array Design via Physics-Guided Decoupling and AI Collaboration

Abstract

In millimeter-wave series-fed array antenna design, simultaneously achieving a low sidelobe level (SLL) and excellent impedance matching over a wide frequency band remains a multi-objective optimization challenge. Traditional parameter global optimization methods based on full-wave electromagnetic simulation incur prohibitively high computational costs when dealing with large-scale arrays. To address this, this paper proposes an efficient design paradigm for array antennas based on physical dimensionality reduction and artificial intelligence (AI) collaborative optimization. The proposed method physically decouples a high-degree-of-freedom $6\times 8$ two-dimensional array into a one-dimensional series-fed array (SFA) and an unequal power divider (PD) network. For each sub-module, a dual-branch fully connected neural network (FCNN) surrogate model is constructed, incorporating electromagnetic prior constraints. Subsequently, the global electromagnetic response of the 2D array is analytically reconstructed using microwave network cascade theory, and efficient multi-objective optimization of core physical dimensions is performed using the particle swarm optimization (PSO) algorithm. Prototype measurement results demonstrate that the E-plane and H-plane SLLs of the optimized antenna are stably suppressed below -17 dB within the 23.7 GHz to 24.7 GHz band, with measured data highly consistent with theoretical predictions. While ensuring physical rigor, this collaborative architecture reduces the traditional full-wave iteration cycle from weeks to mere hours, providing a reliable solution for the agile optimization of complex microwave passive devices.