Rui Liu

A deep-learning-based antenna array synthesis method for varied aperture lengths is proposed in this article. Inspired by Natural Language Processing (NLP), a tri-branched Recurrent Neural Network (RNN) based on Long Short Term Memory (LSTM) layers and “encoder-decoder” structure is designed as the backbone architecture of the proposed model. To further enhance deep models' comprehension of physical scenarios, this article also proposes an improved word embedding method that allows continuous real-valued parameters to be mapped into high-dimensional vectors in a manner similar to individual vocabulary terms, while preserving their original “semantic” relationships. Based the approaches mentioned above, the proposed model is built and operates through an interactive way, similar to a chatbot. Once receiving main lobe direction and array element positions as inputs, it automatically generates amplitude-phase synthesis results that meet specified requirements while ensuring sidelobe levels are minimized to the greatest extent possible. In the numerical experiments, it shows that this method can generate patterns whose main lobe covers the range from 45° to 135° on non-uniform linear arrays with 15-30 isotropic elements, while controlling the side lobe level below 15 dB, effectively improving the directivity of the antenna array. Compared to genetic algorithms, particle swarm optimization and DNN, this method demonstrates significant advantages in both synthesis performance and computational efficiency, significantly reducing the training time while achieving better radiation pattern characteristics.

A dual-branched Convolutional Neural Network integrated with hybrid training criteria is proposed for real-time multi-beam synthesis in planar uniform antenna arrays. For higher speed in dataset generation, pattern superposition method is introduced in this article, which significantly improves data generation speed (~0.7ms per each). With a new hybrid training strategy integrating both “Data-Driven Approximation” and “Physics-Informed Approximation” criteria, flexible training while maintaining model synthesis efficiency is achieved simultaneously. Based on approaches mentioned above, the proposed dual-branched convolutional neural network enables dynamic multi-objective balance between its two branch outputs, accomplishing amplitude-phase synthesis for antenna array design. In the numerical simulation, consistent sidelobe suppression (20~40dB) is demonstrated for dual-beam steering across 10°~20° off-axis angles, with high synthesis speeds (~0.05s per each) on suitable computing platforms. In comparison experiments, the proposed method outperforms typical population-based stochastic optimization algorithms and Deep Neuron Network structure in both synthesis efficiency and pattern regularity. The model's real-time capability is verified for beam reconfiguration scenarios requiring sub-second responses, suggesting a new pathway for physics-embedded deep learning in antenna array synthesis.

A training method based on transfer learning is proposed in this article, which completes the beamforming task of a deep neural network based on ResNet through two stages: "data approximation" and "pattern approximation". It is shown in the numerical simulations that the proposed method achieves better performance in beamforming tasks, compared to the Dolph-Chebyshev method and the Genetic Algorithm. The model can generate complex excitation for arbitrary beam directions within the 0∼45° off-axis angle range and control the sidelobe levels within the 20dB∼40dB range. Furthermore, with the support of suitable computation platform, the synthesis process can be completed in about 0.04 seconds, which meets practical requirements in real-time synthesis, especially in the field of 5G communications.