A recurrent neural network with the residual connection for predicting the pulse characteristics in stress and polarization angle tunable laser

Abstract

The tunable fiber laser admits adjustment of the transmission characteristics of pulses, such as wavelength and width, required in practical applications. Because the available scale and range of tuning parameters are often limited in the experiment, systematic numerical simulations are often used to guide experiments. However, this approach has the disadvantages of consuming long computation times and the necessity of meticulously modeling complex systems. To realize a simpler and, simultaneously, more efficient and accurate approach to the modeling of complex fiber lasers, we have devised a bidirectional long short-term memory recurrent neural network with residual connections (res_BiLSTM RNN) for quick and reliable predictions of the pulse evolution in the fiber lasers, which may be controlled, in particular, by the bend-induced fiber stress and polarization angle. Predictions produced by the neural network for dependences of time-domain and spectral pulse characteristics on the control parameters agree well with experimental results. Compared to traditional numerical simulations, the computational efficiency of res_BiLSTM RNN is improved by two orders of magnitude. In addition, the prediction of the control law in the parameter range, which is difficult to achieve in the experiment, is consistent with results produced by the traditional numerical simulations. The res_BiLSTM RNN scheme can be applied to the real-time monitoring of pulses, providing guidance for the design of tunable fiber lasers, and it may find applications to other settings in nonlinear optics and photonics.