Neural Networks for Real-Time Adaptive Beamforming in Simultaneous Transmit and Receive Digital Phased Arrays: Student Submission

Fully digital phased arrays have become more common as technology advancements have driven down their cost. However, adaptive beamforming techniques can lead to extremely high computational costs when faced with handling the many high-rate data streams created by fully digital arrays. In this work, we focus on the particular challenge of adaptive beamforming for transmit/receive isolation required by Aperture-Level Simultaneous Transmit and Receive (ALSTAR) phased array architectures. We propose that the adaptive transmit and receive beamformers required for a 10-element uniform linear ALSTAR array can be generated by a neural network. In our simulations, the neural network generated beamformers achieved 165.8 dB of effective isotropic isolation at broadside, a gain of 39.7 dB over the non-adaptive beamformers and a loss of only 1.3 dB with respect to the beamformers generated by the ground-truth adaptive beamforming algorithm. This work demonstrates that shallow neural networks are capable of accurately learning the complex input-to-output mapping described by the adaptive beamforming algorithm required for the ALSTAR architecture. We also show that the neural networks generate the required adaptive beamformers to a high precision with a reduced computational complexity (compared to the previously proposed adaptive beamforming algorithm). Because Field-Programmable Gate Array (FPGA) implementations of neural networks exist and can leverage their highly parallelizable structure, we also suggest that neural networks open a pathway to real-time adaptive beamforming in ALSTAR arrays. This is especially true as the FPGAs integrated into phased array platforms grow in size and performance.