Vision Transformer for Beamforming on Phased Array Antennas

Beamforming is not only a major feature to increase the capabilities of 5G and its successors but also it is a key feature of current and future satellite communication systems. While convolutional neural networks (CNN) have already been applied for beamforming, there is no implementation using vision transformers despite their massive advance in image recognition. To the best of our knowledge, we are the first to deploy a vision transformer for beamforming. It successfully predicts the phases of a given antenna pattern for an 8x6 patch antenna array. The final architecture uses multi-head attention as well as convolutional layers for feature extraction resulting in a convolutional vision transformer. The developed model outperforms comparable CNNs, while needing fewer resources making it more suitable for non-terrestrial applications.