Layer-Interaction DeepONet for Modeling Ultrafast Nonlinear Dynamics in Optical Fibers

The propagation of ultrashort pulses in optical fibers plays a crucial role in the design of fiber lasers and optical communications. However, modeling the highly nonlinear dynamics of ultrashort pulses requires extensive numerical simulations. Deep Operator Network (DeepONet), based on the universal operator approximation theorem, provides a novel perspective for modeling ultrashort pulse propagation in optical fibers by approximating nonlinear operators in fibers. However, traditional DeepONet still faces challenges in modeling accuracy when dealing with complex ultrashort pulse dynamics. To address this issue, this paper proposes a novel Layer-Interaction DeepONet, which features a simpler structure and stronger modeling capability for nonlinear fiber dynamics. The superior performance of Layer-Interaction DeepONet is validated through two classical nonlinear fiber dynamics problems: high-order soliton evolution and supercontinuum (SC) generation. The modeling results show that Layer-Interaction DeepONet outperforms in both modeling accuracy and learning speed. Finally, we leverage transfer learning to handle SC scenarios under different nonlinear conditions. Compared to training from scratch, inheriting parameters from a pre-trained model significantly reduces computational cost and sample requirements.