Deep neural network for modeling soliton dynamics in the mode-locked laser
Yin Fang, Haobin Han, Wen-Bo Bo, Wei Liu, Benhai Wang, Yue‐Yue Wang, Chao‐Qing Dai
Abstract
Integrating the information of the first cycle of an optical pulse in a cavity into the input of a neural network, a bidirectional long short-term memory (Bi_LSTM) recurrent neural network (RNN) with an attention mechanism is proposed to predict the dynamics of a soliton from the detuning steady state to the stable mode-locked state. The training and testing are based on two typical nonlinear dynamics: the conventional soliton evolution from various saturation energies and soliton molecule evolution under different group velocity dispersion coefficients of optical fibers. In both cases, the root mean square error (RMSE) for 80% of the test samples is below 15%. In addition, the width of the conventional soliton pulse and the pulse interval of the soliton molecule predicted by the neural network are consistent with the experimental results. These results provide a new insight into the nonlinear dynamics modeling of the ultrafast fiber laser.