Transformer-based long-distance PDM optical fiber channel modeling method

A Transformer-based long-distance polarization division multiplexing (PDM) optical fiber channel modeling method is proposed. The Transformer method's computation time is only 0.5% of that of the Split-step Fourier Method (SSFM). Transformer is a deep learning network. It is based on the self-attention mechanism. Compared to Convolution Neural Network (CNN) and Deep-Learning Neural Network (DNN), Transformer demonstrates superior performance in long-sequence modeling. At a fiber channel length of 1200 km, this method shows good fitting performance for the output optical signal, with a normalized mean square error (MSE) of approximately 0.0007. This method also demonstrates good generalization across different modulation formats and transmission powers. Furthermore, the computational complexity of Transformer is approximately one 1,600th that of the Split-Step Fourier Method.