Predicting the soliton dynamics and system parameters in optical fiber couplers

This work predicts the soliton dynamics and parameter discovery for the coupled nonlinear Schrödinger equations in asymmetric fiber couplers. Using exact one- and two-soliton solutions derived from Hirota’s bilinear method, an extended physics-informed neural networks approach is proposed. This method, which incorporates loss function weight decay and hard constraints, addresses both the inverse and forward problem predictions of the model. Comparisons with analytical solutions demonstrate that this approach effectively predicts soliton transmission and accurately identifies unknown parameters in the inverse problem. Additionally, the study examines how different noise conditions, the number of neuron layers, and dataset selection impact prediction errors.