DNN-based Intelligent Beamforming on a Programmable Metasurface

Abstract

The Programmable Metasurface (PM) can flexibly manipulate electromagnetic waves in real time using loading active devices on the meta-element. Calculating the radiation fields of the PM with complex structures using full-wave simulation software is time-consuming, which results in design efficiency. To accurately and efficiently solve the mapping relationship from coding schemes to radiation fields, an Auto-Measuring System (AMS) of radiation patterns is designed. A few Code-to-Pattern (C-P) data are measured via the AMS. Then, a forward Deep Neural Network (DNN) is proposed, the DNN is trained by the measured data, and an accurate and efficient prediction of C-P is realized. More C-P data are generated based on the proposed forward neural network, and the data are used to train another proposed inverse DNN and realize the accurate prediction of codes when given patterns in real time. This method provides a new alternative scheme for radar beamforming and has application values in intelligent radar beamforming and microwave imaging.

Keywords

• programmable metasurface
• deep neural network
• full connected network
• code-to-pattern
• pattern-to-code