Physics-oriented learning of nonlinear Schrödinger equation: optical fiber loss and dispersion profile identification

In optical fiber communication, system identification (SI) for the nonlinear Schr\"odinger equation (NLSE) has long been studied mainly for fiber nonlinearity compensation (NLC). One recent line of inquiry to combine a behavioral-model approach like digital backpropagation (DBP) and a data-driven approach like neural network (NN). These works are aimed for more NLC gain; however, by directing our attention to the learned parameters in such a SI process, system status information, i.e., optical fiber parameters, will possibly be extracted. Here, we show that the model-based optimization and interpretable nature of the learned parameters in NN-based DBP enable transmission line monitoring, fully extracting the actual in-line NLSE parameter distributions. Specifically, we demonstrate that longitudinal loss and dispersion profiles along a multi-span link can be obtained at once, directly from data-carrying signals without any dedicated analog devices such as optical time-domain reflectometry. We apply the method to a long-haul (~2,080 km) link and various link conditions are tested, including excess loss inserted, different fiber input power, and non-uniform level diagram. The measurement performance is also investigated in terms of measurement range, accuracy, and fiber launch power. These results provide a path toward simplified and automated network management as another application of DBP.