Multi-band systems have demonstrated to be a viable solution to sustain capacity growth required by optical communication systems, thanks to the availability of wide bandwidth amplification technologies, like the Raman amplifier (RA). However, extreme levels of optimization are needed to extract all the potential, requiring super-fast and accurate evaluation of the impact of nonlinear effects. This is a tricky task when the transmission bandwidth is very large, as all fiber parameters becomes frequency dependent and the number of data channels and RA pumps is large. Also, the inter-channel stimulated Raman scattering (ISRS) become impactful. Optimization approaches based on Gaussian Noise (GN) models turn to be very complex, with a consequent slow down of the whole design process. Resorting to the fast GN-based closed-form-models (CFMs), it requires a full spectral and spatial knowledge of the signal power profile along the fiber span. This is particularly computational heavy when backward RA is considered. We propose an approach based on machine learning (ML) and neural networks (NN) to accelerate the process. The method, tested for a super-(C+L) system (12 THz bandwidth) and backward Raman amplification, guarantees a high level of accuracy and a significant speed increase.
Brusin et al. (2025) studied this question.