Learning to predict soliton dynamics in fiber lasers using a recurrent neural network with high accuracy

Accurate and efficient prediction of ultrafast dynamics for a mode-locked fiber laser (MLFL) utilizing deep learning is important for understanding the fundamental physics in ultrafast science and facilitating applications of fiber laser. However, it is still challenging to accurately predict the pulse dynamics of solitons with complex spectral and temporal features using deep learning due to the difficulty of capturing elaborated features in spectral and temporal domains. Herein, we propose and demonstrate a recurrent neural network with bidirectional long short-term memory (Bi-LSTMnet) to precisely predict full-field soliton dynamics in the MLFL. The inferring of the Bi-LSTMnet for 600 round trips of laser is less than 0.1 s, where the result is benchmarked with the ones obtained by the split-step Fourier method (SSFM). Typical prediction results of solitons with high-order Kelly sidebands and soliton molecules with complicated spectral and temporal features consolidate that the Bi-LSTMnet excels in accurately predicting the full-field evolutions of soliton dynamics and generalization capacity for forecasting evolution beyond the training dataset. Our results might facilitate efficient design and optimization of MLFLs. • Bi-LSTM accurately predicts soliton dynamics, surpassing the performance of unidirectional LSTM. • Revealing elaborated features in the full-field evolution of solitons. • Bi-LSTM can predict pulse dynamics beyond the training data range.