Data-driven fiber model based on the deep neural network with multi-head attention mechanism
Yubin Zang, Zhenming Yu, Kun Xu, Minghua Chen, Sigang Yang, Hongwei Chen
Abstract
In this paper, we put forward a data-driven fiber model based on the deep neural network with multi-head attention mechanism. This model, which predicts signal evolution through fiber transmission in optical fiber telecommunications, can have advantages in computation time without losing much accuracy compared with conventional split-step fourier method (SSFM). In contrast with other neural network based models, this model obtains a relatively good balance between prediction accuracy and distance generalization especially in cases where higher bit rate and more complicated modulation formats are adopted. By numerically demonstration, this model can have ability of predicting up to 16-QAM 160Gbps signals with any transmission distances ranging from 0 to 100 km under both circumstances of the signals without or with the noise.