A Machine Learning Approach for Modeling Dynamic Capillary Effect in Supercritical CO2-water Flow

Current theories about capillary pressure and saturation relationship (CPSR) in supercritical CO2-water system are based on measurement results under equilibrium state, which ignores the CPSR dynamic (transient) characteristics, leading to uncertainty in characterizing supercritical CO2-water flow (e.g., underestimating water saturation, overestimating production behaviors of producers, and so on). Consequently, to better understand the supercritical CO2-water flow and reduce uncertainty, it is crucial to study the CPSR transient characteristics. To reduce computational and experimental efforts, a machine learning approach, specifically, an artificial neural network (ANN), was introduced in this work to predict the dynamic capillary coefficient in supercritical CO2-water system by “learning” from available data. For the ANN, the input parameters consist of system temperature, water saturation, petrophysical properties of the porous media (e.g., permeability and porosity), and fluids properties (e.g., viscosities ratio of supercritical CO2 to water, and density ratio of supercritical CO2 to water), while the output parameter is the dynamic capillary coefficient. The results show that the dynamic capillary coefficient can be accurately determined through ANN modeling. In supercritical CO2-water system, dynamic capillary coefficient increases as the system temperature increases. Under a given water saturation, dynamic capillary pressure is larger than the equilibrium capillary pressure. Compared to the conventional CPSR measured under static state, the investigation conducted in this paper provides more valuable insights to the supercritical CO2-water system, which will be helpful for studying two-phase flow system in the context of geological carbon sequestration.