Oil saturation estimation and uncertainty evaluation by modeling-data-driven Gaussian mixture conditional generative adversarial networks

Lacking of training data and uncertainty evaluation of inversion results are challenging problems in deep learning geophysical inversion. In this paper, we propose to combine the Gaussian mixture model (GMM) with the conditional generative adversarial network (cGAN) to solve the problem of prediction evaluation as well as to establish the nonlinear mapping between seismic elastic parameters and oil saturation. Additionally, the synthetic data set generated by the rock physical model is used to train the network to overcome the problem of insufficient training data, and the GMcGAN trained by synthetic data is applied to the real data inversion successfully. Compared with the DNN and CNN, GMcGAN has higher accuracy and can obtain the joint probability density function (PDF) of oil saturation. Compared with the well logging data training, the modeling data-driven GMcGAN can avoid the empirical risk and reduce the uncertainty of prediction, the results are more consistent with the actual drilling results.