Fast and accurate waveform modeling based on sequence-to-sequence framework for multi-channel and high-rate optical fiber transmission.

Abstract

We propose a sequence-to-sequence (Seq2Seq) framework integrated with a feature decouple distributed (FDD) method for fast and accurate channel waveform modeling in multi-channel, high-rate wavelength-division multiplexing (WDM) optical fiber transmission. This framework enables the simultaneous prediction of multiple output symbols in a single inference, dramatically reducing the repeated calculation of adjacent padding symbols and achieving a significant reduction in time complexity compared to the traditional split-step Fourier method (SSFM). Additionally, transfer learning is leveraged to streamline the training process and improve the accuracy of the Seq2Seq architecture. In a 40-channel, 140 GBaud WDM system, Seq2Seq-FDD reduces computation time to a mere 0.22% of that required by the variable step size SSFM. In a five-channel configuration, Seq2Seq-FDD achieves an 85.5% improvement in NMSE over simplified FDD-Co-LSTM and a 99.88% reduction in computation time compared to vanilla-FDD. This framework provides a highly efficient solution for waveform modeling in multi-channel, high-rate WDM systems.