High-Speed Channel Modeling Using Transformer Network

Abstract

This article proposes a high-speed channel transformer (HSCT) network module to model the S-parameter of traces. To deal with length variation, RLGC matrices are modeled using transformer networks and converted by an RLGC-to-S layer. For validation, the accuracy of predicted RLGC and S-parameters is compared with conventional simulators. As a result, error rates of estimated magnitude and phase achieve 0.76% and 3.62%, respectively. The pretrained HSCT module is verified to be scalable in terms of the number of frequency points, without redefining weights and additional training. In addition, a precise analysis of the RLGC data distributions depending on preprocessing techniques is performed, hence, to ensure the reproducibility of the proposed method.