It is of great significance to reduce the uncertainty of seismic phase analysis results by combining multiple deep learning algorithms to mine hidden and useful information in seismic data, and to achieve mutual complementarity and optimization. Therefore, a method and process of deep learning seismic phase analysis from label training to data mining to optimization were proposed. Firstly, waveform classification is performed by the SOM of the self-organizing mapping network diagram, which provides representative training data for supervised learning. Then, the convolutional neural network CNN and the circulating neural network RNN are used for seismic phase analysis, and the predicted seismic phase analysis results are input to the generative adversarial neural network GAN for optimization between algorithms and uncertainty analysis of operation results, and finally the optimal results are given based on actual data analysis. The method and practical process of SOM+CNN/RNN+GAN combined supervised and unsupervised deep learning seismic facies analysis are proposed and realized, and it is proved that the method improves the reliability and effect of seismic facies analysis and oil and gas reservoir prediction results through the practical application of oil and gas prediction in river channel sand reservoir reservoirs in the study area.