Multi-input neural channel waveform model for optical fiber WDM transmission based on Volterra series transfer function

Accurate waveform modeling of optical fiber channels is essential for the design, optimization, and management of wavelength-division multiplexing (WDM) systems in optical communication networks. To address the computational inefficiencies of the traditional split-step Fourier method (SSFM), deep learning has achieved significant advancements in this field. However, current deep learning based channel models take only transmitted signals as inputs, achieving good generalization for system parameters that can be derived from the signal waveform. For system parameters such as baud rate, dispersion and nonlinearity coefficients that cannot be extracted from the waveform, any variation in these parameters necessitate retraining the model, thereby limiting its flexibility. Here, we build upon the existing physics-based Volterra series transfer function algorithm and employ neural network parameterization in the frequency domain (NN-VS) to achieve high-accuracy and robust generalization modeling of WDM channels with support for multi-parameter inputs. We evaluated the performance of NN-VS in simulations of a 40-channel 600 km and a 5-channel 1200 km WDM system. Under various baud rates, dispersion coefficients, and nonlinearity coefficients, the proposed NN-VS scheme achieved an average Q-factor error of less than 0.15 dB at the optimal launch power. Furthermore, NN-VS demonstrates superior computational efficiency compared to SSFM, achieving transmission in a 40-channel WDM scenario with less than 2% of the real multiplications while delivering millisecond-scale runtime on a GPU.