Integration of Convolutional Neural Networks and Autoencoding for Generating Reconfigurable Intelligent Surfaces

This paper presents a method utilizing convolutional neural networks (CNN) and autoencoding for generating a reconfigurable intelligent surface (RIS) based on information like beam angles and radiation patterns. It reports how to solve a complex problem in communications with artificial intelligence through requirements analysis and designing of models. Generating a RIS for meeting the strict requirements of beam-steering is challenging since beamforming depends on the elements on the surface which can be arranged in a vast amount of combinations. A deep learning approach has been considered, but finding a suitable model is also a difficult task. This paper introduces a method of generating a small-size RIS and gradually introduces different design tactics to apply the model to solve the problem of larger RIS which is able to handle different incoming and outgoing beam angels. The design rationales of architecture and functions are introduced in the method and the corresponding experiments so that the practitioners can follow our method to construct an RIS. We employed various architectural improvement approaches in a sequential manner, utilizing a step-by-step method to enhance our neural network models. These modifications were carried out to demonstrate the model's capability in feature extraction and its increased generalization capacity. This was substantiated through the use of two distinct sizes of RIS, showcasing a notable enhancement in both feature-capturing ability and overall generalization performance.