Phase information is an important parameter in the wave function of a Bose-Einstein condensate in an optical lattice. However, in experiments, the phase information of the wave function cannot be obtained directly from the atom distribution in momentum space by absorption imaging or in-situ imaging. Thus, a deep learning network model was developed to study the influence of the phase distribution of a Bose-Einstein condensate on the atom distribution in momentum space. Thirty-two thousand data sets obtained by theoretical calculations were used as training and verification sets. Based on the analysis of the phase characteristics and momentum space of the wave function, a method for predicting the momentum of supercooled atoms in an optical lattice was developed using a convolutional recurrent neural network model. After the model verification, a difference between the model training and Schrodinger equation results is 1.76, which is 83% less than the average error of a back propagation neural network. Our approach provides a new application of machine learning in the field of physics.