Modeling of 2-phase flow in porous media using physics-informed neural network

Abstract

Oil and gas are valuable resources and make a great contribution to the national economy. However, this resource is often located thousands of meters underground. Therefore, oil and gas production and exploitation conditions cannot be measured directly but must be through simulations. Specifically, the models here simulate the behavior of fluids in oil and gas reservoirs to predict and optimize oil and gas exploitation. Besides traditional numerical simulation models, artificial neural networks are a good alternative. This study investigates and presents the advantages as well as disadvantages of the physics-informed neural network (PINNs) approach in the modeling of a simple oil production process by water flooding method with the Buckley-Leverett theory. The results from PINNs in this study agree quite well with the analytical solutions of the differential equation to be solved. The results from this work also show that PINNs can be applied to problems in which the unknowns have stiff changes with respect to space/time and that PINNs are absolutely suitable for problems where data availability is limited such as those for oilfield development/production planning. Furthermore, PINNs trained models are reliable and able to be utilized for long term production forecasts thanks to the consideration of physics information during the training of the network. This shows that PINNs would be highly applicable in the modeling of two-phase flows in porous media.