Abstract

In modern optical transmission systems, accurate and reliable simulation of optical fiber plays a crucial role in link system design, transmission performance prediction, signal impairment analysis, and compensation algorithm verification. Machine learning (ML) has contributed to novel developments in this field, successfully solving the problem of slow computation in the traditional split-step Fourier method. However, many ML-based optical channel models sacrifice flexibility and can only be applied to few scenarios such as single-channel system or fixed-loaded wavelength-division multiplexing (WDM) system limited to C band. Deep operator network (DeepONet) as a powerful neural operator was proposed as an alternative solution for multiple application scenarios. In this study, we proposed DeepONet-based waveform-level simulation techniques for optical fiber communication, including the classical DeepONet for single channel modeling, the stacked DeepONet for fully-loaded C+L-band WDM system, and the loading condition aware DeepONet for randomly-loaded system. To evaluate those models, we compared DeepONet with other ML-based models and studied multiple types of signals including varying signal powers, modulation formats, transmission distances, and loading conditions. The better performance on flexibility and generalization were achieved, and the mean square errors (MSEs) were lower than $10^{-3}$ under all application scenarios, which demonstrates the feasibility of DeepONet for waveform-level simulation in wideband WDM systems.