Direct prediction of relative permeability curve from 3D digital rock images based on deep learning approaches

Abstract

The relative permeability curve is one of the key features to evaluate the flow property of a porous medium, which is important in petroleum and gas industry, yet it is not easy to obtain. In this study, we proposed a new intelligent way to predict the relative permeability curves directly from 3D digital rock images, compared with other existing artificial intelligence (AI) methods that usually used indirect geometrical parameters as inputs. The inputs of our AI model are the digital rock images and fluid/rock physical properties such as wettability and interfacial tension; the outputs (relative permeability curves) are obtained by an improved pore-network model. We test three deep learning methods (CNN, ConvLSTM-FC, and ConvLSTM-CNN) and compare their accuracy in the predictions for both sandstone and volcanic rock types. The results show that the overall prediction accuracy of the deep hybrid ConvLSTM-CNN method is the highest, reaching 95%. The well-trained model is further applied to successfully predict the effects of wettability and interfacial tension on the relative permeability curves of a Berea sandstone.