Graph Neural Network Assisted S-Parameter Inference and Control-Word Generation of Terahertz Reconfigurable Intelligent Surface

Zihan Ning, Tong Sun, Jizhao Li, Yanfei Ren, Chenjia Xie, Li Du, Lianggong Wen, Yuan Du

Abstract: For large-scale terahertz (THz) reconfigurable intelligent surfaces (RIS), each unit element is supposed to be controlled by an independent voltage source. This leads to a huge solution space for both key metrics inference and control-word generation. In this paper, we propose a fast AI-assisted control-word generation scheme to reduce the computation cost of Electro-Magnetic (EM) simulations (Forward Model) and to accelerate the iteration process for control word search (Inverse Model). The results demonstrate that the Forward Model can predict the S-parameters between 100GHz and 800GHz with a minimal mean absolute error (MAE) of 0.69dB. Our method is more than 180 times faster than traditional full-wave simulation methods without training time. Additionally, the Inverse Model can generate demanded control words within 1.5 dB error requirement in less than 200 iterations.