Learning to shape beams: Using a neural network to control a beamforming antenna

The field of reconfigurable intelligent surfaces (RIS) has gained significant traction in recent years in the wireless communications domain, owing to the ability to dynamically reconfigure surfaces to change their electromagnetic radiance patterns in real-time. In this work, we propose utilizing a novel deep learning model that innovatively employs only the parameters of each signal or beam as input, eliminating the need for the entire one-dimensional signal or its diffusion map (two-dimensional information). This approach enhances efficiency and reduces the overall complexity of the model, drastically reducing network size and enabling its implementation on low-cost devices. Furthermore, to enhance training effectiveness, the learning model attempts to estimate the discrete cosine transform applied to the output matrix rather than the raw matrix, significantly improving the achieved accuracy. This scheme is validated on a 1-bit programmable metasurface of size 10 × 10, achieving an accuracy close to 95% using a K-fold methodology with K=10. • Development of a compact deep decoder to translate beam parameters to RIS pattterns. • Using DCT transform as the output of the network, enhancing the model’s performance. • Errors of less than 4% are achieved in estimating the RIS pattern for a shaped beam.