Predicting Ultrafast Soliton Evolution via a Parameter‐Driven Neural Network

ABSTRACT Numerical approaches based on the nonlinear Schrödinger equation are computationally intensive and inefficient to study soliton formation and evolution in fiber lasers, while existing temporal neural networks can accelerate evolution prediction but remain constrained by their dependence on simulations during prediction. We propose a parameter‐driven Transformer–Temporal Network framework to end‐to‐end predict soliton dynamics directly from laser‐cavity parameters. In this framework, the Transformer module predicts the initial pulse from the cavity parameters, and the temporal network then models the entire soliton evolution, thereby eliminating the need for simulation‐derived initialization. Using datasets generated from coupled‐NLSE simulations and experimental validation in a mode‐locked fiber laser, the model accurately captures the full evolution of bound‐state solitons from nonstationary to steady states, and reproduces key dynamical phenomena such as the periodic spectral parity switching of bound solitons. The proposed approach maintains high predictive accuracy and strong consistency with experimental observations, while the Transformer module exhibits stable generalization under multi‐parameter perturbations. This work realizes an end‐to‐end temporal neural network prediction from laser‐cavity parameters to complete soliton evolution, achieving a prediction speed approximately 150 times faster than simulations and thereby extending the independent applicability of neural networks in fiber‐laser physics.