Conditional Fourier Neural Operator for Prediction of Ultrafast Nonlinear Dynamics in Optical Fibers

Ultrafast nonlinear dynamics in optical fibers as the cornerstone of ultrashort laser pulse transmission and control is significant for the rapid development of mode-locked fiber lasers, fiber amplifiers, and fiber communication technologies. Solving the nonlinear Schrödinger equation using traditional numerical methods is very time-consuming, which greatly limits the design of experiments and real-time optimization. Deep learning technology has been introduced to predict ultrafast nonlinear dynamics in optical fibers and has achieved significant breakthroughs. However, existing deep learning-based methods mainly focus on the effective prediction of single scenarios. Moreover, the prediction accuracy of deep learning-based methods for ultrafast nonlinear dynamics in optical fibers needs to be further improved. In this study, a conditional Fourier neural operator (FNO) is proposed to predict the ultrafast nonlinear dynamics of multiple typical propagation scenarios in optical fibers accurately. A total of six independent tasks are completed by a single model of conditional FNO. The results show that the conditional FNO can predict the ultrafast nonlinear dynamics of different input pulses propagating under different fiber parameter settings with extremely high accuracy and strong generalization ability. In terms of accuracy metrics, the conditional FNO achieves state-of-the-art results. Strong generalization ability is the inherent advantage of neural operators. Our work provides an advanced and powerful generalized model for predicting ultrafast nonlinear dynamics, offering important insights into the widespread applications of deep learning in ultrafast optics.