Predicting nonlinear reshaping of periodic signals in optical fibre with a neural network

The accumulation of nonlinear effects in an optical fibre is often seen as a source of significant impairment for the propagating light signals, but the same effects, when properly managed, can provide a remarkable tool to tailor the temporal and spectral content of the signals. Indeed, depending on the regime of dispersion of the fibre and the frequency chirp, an initial pulse can be significantly expanded or compressed in the time or frequency domain, or it can be reshaped into advanced temporal waveforms such as parabolic, rectangular and triangular shapes [1]. Yet, due to the typically wide range of degrees of freedom involved, predicting the behaviour of nonlinear pulse shaping by numerical integration of the nonlinear Schrödinger equation (NLSE) or its extensions may be computationally demanding, especially when dealing with inverse-mapping problems. Recently, we have successfully introduced the use of the machine-learning (ML) method of artificial neural networks (NNs) as an efficient tool for complementing or substituting the NLSE in the modelling of nonlinear pulse shaping [2-5] or for predicting the generation of optical supercontinua [6, 7]. Fibre nonlinearity does not only affect the propagation of ultrashort pulses.