Abstract
A prior-knowledge-guided transformer-model-based synthesis method is proposed for extrapolating active S-parameters and full S-parameter matrices in linear antenna arrays, accounting for mutual coupling effects. The method predicts both the active S-parameters and full S-parameter matrices of larger arrays using training dataset generated from smaller arrays. By formulating the task as a sequence-to-sequence transformation, a transformer-model-based neural network, as one type of machine learning methods, learns the mapping between excitations and reflections at each feeding port of large-scale antenna arrays. Prior knowledge−including features of the antenna arrays and mechanisms of transformer models−is incorporated to enhance model performance and physical consistency. In full-wave simulations, the proposed transformer model is trained on simulated data from linear arrays with fewer than 20 elements and then used to extrapolate the active S-parameter and full S-parameter matrix for a 30-element array. Experimental validation using fabricated printed dipole arrays further demonstrates the accuracy and robustness of the proposed method, as supported by both simulation and measurement results.
Funding Information
  • Adelaide Research and Innovation, University of Adelaide
  • National University of Singapore (ARIS-RP1)
  • Ministry of Education and Research, Romania (MOE-000742-01/T2EP50124-0047)

This publication has 47 references indexed in Scilit: