Prior Knowledge-Incorporated Deep Learning for Antenna Array Design Using Method of Maximum Power Transmission Efficiency

This paper proposes an innovative prior-knowledge-incorporated deep learning method for the design of both near-and far-field antenna arrays. By integrating the method of maximum power transmission efficiency as physical prior knowledge, the neural network is guided to predict the scattering parameters between virtual and transmit antennas, rather than directly mapping excitations to beam patterns. This shift in the design objective provides a more flexible framework for beam pattern control. Furthermore, this integration yields a compact dataset of highly efficient samples, enabling the model to achieve high accuracy even with small-sample training. A 4 × 4 microstrip array with different near-and far-field functions is designed for demonstrating the proposed method. A 6-layer full-connection deep learning model is applied in near-field scattering parameters prediction and extended to far-field scattering parameter prediction using meta-learning and transfer learning. The simulated and measured results agree well with the predicting results, demonstrating superior performance in applications including near-field focusing, beam steering, multi-beam generation, and flat-top beam shaping. This work thus provides a crucial technical pathway for large-scale antenna array design in wireless communication and radar systems.