Neural Networks for predicting optical pulse propagation through highly nonlinear fibers

Due to increase in demand of the optical fiber communication system there is a special emphasis on diagnosing ultrashort pulses. The linear and nonlinear distortions introduced during transmission gives rise to wide variety of wave dynamics. The conventional signal processing techniques being used for characterising these pulses are computationally inefficient. Since machine learning has shown improvement compared to other analytical methods, we present a comparative study of different neural network (NN) architectures to predict the output pulse profile after transmission through highly nonlinear and dispersive fibers. The trained network has the ability to learn the mapping from a set of input and output pulses for the case of both known and unknown fibers. Since each NN has its own advantages and disadvantages, we to the best of our knowledge, present a comprehensive analysis of six different NN architectures (i) fully connected NN (FCNN), (ii) cascade forward NN (CaNN), (iii) Convolutional NN (CNN), (iv) long short term memory network (LSTM), (v) bidirectional LSTM (BiLSTM) and (vi) gated recurrent unit (GRU) for the first time.