A Deep Learning-based Approach for Inverse Design of Reconfigurable Metasurfaces

Reconfigurable metasurfaces (RMTSs) have attracted increasing attention for various practical applications thanks to their intriguing ability to manipulate electromagnetic (EM) waves dynamically. However, conventional modelling and optimization methodologies of RMTSs can be ad-hoc and time-consuming, which restricts their development. This paper proposes a novel inverse design methodology based on deep learning (DL) techniques to train an inverse artificial neural network (ANN) to predict the corresponding physical dimensions of RMTSs with a given EM response in real time with high predictive accuracy. This methodology is modified from the conventional forward and inverse design by introducing performance indices including resonant frequency, max amplitude, and 3dB bandwidth of the reflection coefficient S11’s first resonant frequency band as the inverse ANN model’s input and the criteria to separate the datasets. Based on the efficient and accurate inverse ANN model, RMTSs can be easily optimized to fit various applications in 5G wireless communications.