C+L Band Gain Design in Few-mode Fibers Using Raman Amplification and Machine Learning

Abstract

A machine learning technique was recently proposed to optimize the gain of a multi-pump single-mode Raman amplifier, using neural networks to approximate the function that maps a given gain profile to the corresponding set of pump powers and wavelengths by training them on synthetic datasets of Raman gains. This method was then extended to FMFs, showing good results in terms of gain flatness and mode-dependent gain, but limited to the C band only. In this paper, we show that the design choice of the dataset generation phase can impact the quality of the neural network predictions, and propose different techniques to improve their accuracy. We present improved results on both flat and tilted gain profiles on the entire C+L band using a few-mode fiber supporting the LP01 and LP11 mode groups and using 8 Raman pumps.