Application of machine learning in predicting flow and transport in porous media

Micro-CT imaging and pore-scale modelling have developed rapidly over the last decade by bridging the disciplines of geology, reservoir engineering, image processing, and computational fluid dynamics. They have provided new pathways for understating complex transport phenomena in heterogeneous geological formations. However, direct simulation of flow in these complex three-dimensional geometries can be difficult and time-consuming. Machine learning and Convolutional Neural Networks (CNN), as a part of the broader field of Artificial Intelligence (AI), can be integrated into the framework of pore-scale modelling. We propose a neural network architecture that considers features of the rock geometry as well as the conservation of mass and predicts the velocity distribution on the images. The method can be applied to two or three-dimensional rock images. The reliability of the flow prediction is studied by comparing the predicted permeability versus ground truth values as a bulk measure. Then, we study the accuracy of solute transport modelling. The results show that velocity fields obtained by CNN can have a considerable degree of error and are not suitable for accurate transport simulations. Finally, challenges and opportunities for the development of machine-learning approaches in porous media applications will be discussed.