A Neural Network Approach to Predict Oil Volume Production Considering Porous Media Images

Abstract

Two-phase flow through porous media is a complex problem since changes in the porous media geometry can cause a significant impact on the fluid flow. The geometry change affects the system permeability of each phase, which is the coefficient linking the flow rate response to an imposed pressure differential. The phase permeability is also a function of the system saturation (ratio between the phases). This flow is costly to predict through direct simulation methods. Moreover, it is tricky to compute the two-phase flow behavior by simpler methods due to the high variability associated with geometry and fluid properties. In this work, we take 2D porous media images as input. We consider the porous media to be initially fully saturated with oil. Water is injected on a porous media side to displace the oil to the opposite extreme. We assembled a neural network system using the DeepONet concept to predict the volume of oil produced over time. The networks demonstrate to calculate the total oil produced presenting a small relative error.