Intelligent Method of Saturation Prediction Using Multiscale Data from Digital Core Image, Core Sample, and Well Logs

Summary Tight sandstone reservoirs are characterized by poor porosity and permeability, making fluid logging responses unclear. As a result, the accuracy of saturation prediction models established based on logging petrophysics is low, making it difficult to meet production demands. In this study, we propose an intelligent saturation prediction method that combines data augmentation and multiscale data integration using deep neural networks (DNNs). The method integrates digital cores, cores, and well logging data to fully utilize macroscopic reservoir information at the logging scale and microscopic fluid properties at the pore scale. To augment the data set, new digital cores are reconstructed using generative adversarial networks (GANs), expanding the training data set for the first time. Oil-water displacement simulations are then performed on the digital cores using mathematical morphology, with different initial saturations assigned to each digital core, further expanding the data set. Data from all three scales—pore, core, and well logging—are integrated to establish the training data set, and the best DNN model is selected for saturation prediction. This approach addresses the challenge of sufficient saturation labels from production testing or sealed coring. For the first time, it uses data from three different scales for saturation prediction. The number of saturation labels has increased by approximately 21 times, and the average absolute error of the DNN model on the testing set is less than 4%, demonstrating the model’s accuracy. A case study shows that the saturation prediction results are in good agreement with those from sealed coring.