Efficient 2-D Pattern Synthesis for Planar Series-Fed Microstrip Antenna Arrays Using Physics-Informed Graph Convolutional Networks

Compared to parallel-fed antennas, series-fed microstrip antennas (SFMAs) exhibit strong coupling among the series-fed elements, leading to interdependent amplitude and phase distributions. This inherent coupling introduces additional design constraints and increases the computational complexity of SFMA synthesis. To address these challenges, a physics-informed graph convolutional network (PIGCN) framework is proposed to optimize the excitations of SFMAs for specific pattern objectives. By projecting the physical topology of the SFMA onto graph structures, mutual couplings are inherently embedded in the neural network. This projection allows information propagation along transmission paths, eliminating the need for extra learning of interelement dependencies. Furthermore, an attention mechanism calculates the excitation adjustments based on correlation coefficients between pattern errors and element excitations. This integration improves sensitivity to dominant error sources and accelerates convergence. Two numerical experiments for different synthesis scenarios demonstrate that, for SFMAs, the proposed framework offers superior accuracy, efficiency, and flexibility compared to the particle swarm optimization (PSO) and the PINN-based method.