An efficient pattern prediction method of array antenna using graph neural network

Phased array antennas are widely used in modern wireless systems, where rapid and accurate radiation pattern prediction is essential for array analysis, design, and optimization. However, pattern prediction remains challenging due to diverse array configurations and mutual coupling effects. This paper proposes an efficient graph neural network-based method for array pattern prediction. The proposed approach predicts active element patterns to includes coupling effects. By representing array elements as graph nodes and their spatial interactions as graph edges, multi-layer graph convolutions are used to aggregate neighborhood information and learn coupling-induced variations in element patterns. A direction-weighted and symmetry-aware loss function is further introduced to improve accuracy and symmetry. Validation is conducted on two fabricated and measured arrays, where the proposed method achieves median element-level prediction errors of 0.3-1.2 dB, with 96.5% of errors below 3 dB. The average array-level prediction errors is significantly reduced compared with the case where coupling is ignored. The scalability is further verified on a numerically simulated 64-element array, for which an average prediction error of 1.57 dB is obtained. In addition, the proposed model maintains microsecond-level inference speed, providing an efficient and practical complement to repeated full-wave simulations for array antenna pattern analysis.