Predicting the Evolution of the Supercontinuum Generation With CNN-LSTM Model

Abstract

We propose a hybrid deep learning model, namely convolutional neural network–long short-term memory (CNN-LSTM) approach to investigate the evolution of the supercontinuum (SC) generation numerically. The hybrid model can use the CNN model to extract and map the local features of the sequence, followed by the LSTM to predict the overall trend of the SC generation. With the trained model by learning the propagation dynamics of the generalized nonlinear Schrödinger equation, the consistent outcome for the neural network predictions and numerical solutions has been obtained. The combined neural network can effectively solve the complex nonlinear propagation problems and maintain high accuracy compared with the LSTM, GRU neural networks for different incident power. The hybrid approach can facilitate the design and optimization of the spectral or temporal intensity distribution of SC generation, and may offer guidance for designing SC source for specific applications.