Abstract:Programmable metasurface (PM) exhibits powerful capabilities to manipulate electromagnetic (EM) waves with controlled active components loaded on the sub wavelength elements. However, for electrically large metasurface with complex structures, it is difficult to implement the fast mapping from the codes, i.e., states of the active components, to radiation patterns and vice versa. In this article, artificial neural networks are employed to realize code-to-pattern (C-P) and pattern-to-code (P-C) mapping accurately and efficiently. As for the C-P mapping, a novel physics-inspired neural network (PINN), which is primarily inspired by the discrete dipole approximation (DDA) method, is proposed. The PINN is physically interpretable and shows a strong few-shot learning ability. The average error between the patterns predicted by the PINN and measured patterns is as low as 2.31 dB. For the P-C mapping, a deep neural network (DNN) is proposed with PINN as its teacher network, i.e., PINN is used to generate more radiation patterns to train the DNN. The average accuracy of codes prediction is higher than 98.4%. Finally, an intelligent beamforming scheme is implemented by combining the PINN and DNN. For the desired patterns, the required codes could be accurately calculated by the DNN in real time. Then, the synthesized patterns of the required codes could be achieved by the PINN to compare with the desired patterns. The proposed scheme is a first step toward practical applications of PM in the fields of sensing and communication.

Physics-driven unsupervised deep learning network for programmable metasurface-based beamforming

Programmable Metasurface-Based Beam Forming by Physics-Driven Deep Learning Network.

Dual-Beam Forming Based on Physics-Driven Deep Learning Method for Programmable Metasurface

Machine-learning Reprogrammable Metasurface Imager.

Coding Programmable Metasurfaces Based on Deep Learning Techniques.

Programmable Transmission Metasurface Scattering Control under Obstacles Based on Deep Learning

Intelligent Beamforming Via Physics-Inspired Neural Networks on Programmable Metasurface

On the Application of Deep Learning in Modeling Metasurface

Space-Time-Modulated Wideband Radiation-Type Programmable Metasurface for Low Sidelobe Beamforming

Experiment-based deep learning approach for power allocation with a programmable metasurface

Programmable Metasurface Intelligent Beamforming

Flexible Beamforming Using Transmission-Type Coding Metasurface

Deep Learning Based Phase-Gradient Metasurface Design with Near-Field Coupling Considered

On-demand Doppler-offset Beamforming with Intelligent Spatiotemporal Metasurfaces

Machine Learning Design Metasurface for Arbitrary Beam Generation and Regulation

Deep Learning of Reflection Phase Predection for Arbitrary Coding Metasurface Atoms

Deep learning estimation of complex reverberant wave fields by a programmable metasurface

A Deep Learning Approach for Objective-Driven All-Dielectric Metasurface Design

Programmable Coding Metasurface for Dual-Band Independent Real-Time Beam Control

Physics-Informed Inverse Design of Programmable Metasurfaces

Coding Metasurface for Shaping Beams Through Software-Based Approach