Compositional reservoir simulation is a time bound activity demanding complex physics. We review the advantages of machine learning in complex compositional reservoir simulations to determine fluid properties, such as critical temperature and saturation pressure. A machine learning approach to predict critical temperatures during simulation based on the Heidemann-Khalil method is implemented, resulting in more accurate results with lower computational cost, outperforming the standard method and improving performance on a giant field model with compositional gradient and miscible gas injection.

The fluid column grades from black oil to gas condensate; accurate phase behavior and miscibility modelling involves a significant number of components. This makes simulation performance one of the biggest challenges. Critical temperature of the mixture is commonly used to determine phase state (phase labeling), a crucial process of reservoir simulation. Mislabeling can result in incorrect physics and convergence issues, particularly in cells with gas-oil displacement. Simulators generally use the Li correlation to calculate the pseudo-critical temperature from the weighted average of component critical temperatures, leading to inaccuracies. The Heidemann-Khalil method is computationally costly, proportional to the cube of the number of components, prohibiting its use for complex compositional simulation models. We use machine learning to efficiently incorporate it into simulation.

By using the machine learning approach, neural networks are trained based on a combination of feeds to reproduce the Heidemann-Khalil method with great precision. The accuracy of critical temperature and other fluid property determinations is thus improved, with machine learning ensuring a very low computational cost. Convergence is also improved. With the traditional Li method, especially at the beginning of gas injection, we faced numerical difficulties, and the runtime was slowed down. By implementing the machine learning based method, the convergence is smooth through the entire gas injection cycle, leading to a reduction of total iteration counts. We experience an overall four times speed-up of the simulation model, which greatly enhances the usage of this model in simulation studies.

The use of machine learning methods to replace physics in the simulator is an evolving area. By showing a field example of a successful application that improves both accuracy and performance, we contribute to fostering research into new possibilities where physics-informed models will enhance simulation studies.

You can access this article if you purchase or spend a download.