Enhancing generalization in neural network-based waveform-level channel modeling for optical fiber transmission...

Fast and accurate optical fiber channel waveform modeling is essential for the evaluation of optical communication systems. Traditional modeling methods, such as the split-step Fourier method (SSFM), suffer from significant computational complexity. In contrast, neural network (NN)-based approaches not only achieve comparable accuracy but also significantly reduce computational burden. Enhancing the generalization capability of NNs is crucial to facilitate flexible system design and optimization across varying system parameters. In this work, we introduce a novel parameter encoding structure, which significantly improves NN generalization by pre-encoding system parameters. In the generalized scenario, with a wide range of launch power from -2 to 7 dBm and arbitrary transmission distances, the parameter encoding structure improves waveform modeling accuracy by 49.9% and 69.7% compared to the non-encoded scheme. Notably, for the first time, we develop a single NN that generalizes across multiple system parameters-modulation format, symbol rate, WDM channel space, phase noise, and frequency offset of lasers, launch power, accumulated chromatic dispersion, span length, and total transmission distance-simultaneously. This enhanced generalized NN holds significant potential for the design and optimization of optical transmission systems.