Electrostatic Potential Distribution Prediction Using Convolutional Neural Networks

Digital images of rocks obtained using high-resolution, micro-computed X-ray tomography (micro-CT) have become an increasingly popular aid in studying pore-scale processes. Petrophysical properties such as permeability, porosity, and fluid saturation are important factors in estimating the amount of hydrocarbon in place and determining the mechanisms of fluid flow through porous media. One of the most common methods of delineating hydrocarbon saturation in rock is measuring the electrical resistivity, a quantification of how strongly a material resists electric current. Using micro-CT images and appropriately identifying the grain and fluid phase distributions, one can numerically calculate the resistivity and electric potential distribution through an arbitrary porous medium. However, wholly deterministic calculations of pore-scale electrical properties require hours to days of simulation using high performance computing and cannot be easily generalized to complex, heterogeneous reservoir rocks.

Here we show the feasibility of applying deep learning techniques to predict the distribution of electric potential in three-dimensional porous media. A convolutional neural network (CNN) is set up to compute geometrical information and extract spatial relationships between the porous medium morphology and the potential field. The CNN is trained using direct pore-scale simulation in a granular medium. The results show an accurate prediction of the potential field through various granular and slightly consolidated porous media, both model and natural, within a few minutes. Using CNNs to accelerate the calculation of a potential field can lead to automating the estimation of hydrocarbon saturation in real time.