Key points are not available for this paper at this time.
Stimulated Raman scattering (SRS)-induced inter-channel power transfer introduces global cross-channel coupling in the amplitude domain, causing power fluctuations across the spectrum that further interact with accumulated Kerr nonlinearities. The C + L-band system modeling and generalization become challenging under various nonlinear effects, especially with power pre-tilt. In this Letter, we propose an enhanced self-attention-assisted multi-channel waveform modeling to achieve efficient and accurate modeling with SRS and non-flat launch power. To capture these cross-channel and long-range dependencies, we apply rotary positional encoding to the query (Q) and key (K) matrices in the attention mechanism. Benefitting from enhanced self-attention, we realize waveform modeling with strong generalization ability across different non-flat launch power profiles and transmission distances in ultra-wideband (UWB) wavelength-division multiplexing (WDM) systems. We compare the split-step Fourier method (SSFM) with the proposed method over a 10-span link at the optimal launch power, and the Q-factor differs from SSFM by only 0.31 dB. In a 5-span scenario with pronounced nonlinear effects, our method reduces runtime by 99.4% while maintaining a Q-factor deviation of just 0.24 dB from SSFM.
Wang et al. (Mon,) studied this question.