Recent advances in deep learning RF applications and metamaterial characteristics prediction

Abstract Over the past decade, deep learning (DL) and its use for radio frequency (RF) engineering and metamaterials has become increasingly important. The previous techniques used in RF processing and the design of metamaterials are limited to their high computation requirements, complex analytical design, and poor ability to model nonlinearities. DL offers a new class of tools to develop data-based models of RF signals and design RF systems through the process of Inverse Design, and allows for real-time analysis with very high analytical accuracy. The authors present a comprehensive discussion of prominent Convolutional Neural Networks, Recurrent Neural Networks, and Generative Adversarial Networks, as well as other recent developments in DL applications for RF Engineering, including spectrum sensing, channel estimation, signal classification, and waveform generation. In addition, this paper also describes how DL models can predict the electromagnetic properties of metamaterials, which allows for inverse design and rapid prototyping using Autoencoders and Surrogate Models. Finally, this paper discusses the simulation tools, essential datasets, and performance evaluation metrics that were used to compare the performance of DL to traditional methods. While the initial results are promising, there are still issues with the current state of the art including limited data, low interpretability of the model, and poor generalizability across frequency bands and geometries. Future research directions will be provided in this paper, including the implementation of real-time intelligent metasurfaces and transfer learning using Physics-Informed Neural Networks (PINNs).