Physics-Informed Machine Learning Applied to Complex Compositional Model in a Giant Field

This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper IPTC 23730,“Physics-Informed Machine-Learning Application to Complex Compositional Model in a Giant Field,” by Guido Bascialla, SPE, ADNOC, and Coriolan Rat and Soham Sheth, SLB, et al. The paper has not been peer reviewed. Copyright 2024 International Petroleum Technology Conference. Compositional reservoir simulation is a time-intensive activity demanding complex physics. In the complete paper, the authors review the advantages of machine learning (ML) in complex compositional reservoir simulations to determine fluid properties such as critical temperature and saturation pressure. An ML 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 case study refers to a giant offshore carbonate field composed of multiple reservoirs. Production is currently in a rampup phase; crestal miscible hydrocarbon gas injection was implemented soon after startup. The availability of high-potential gas producers as a source of makeup gas and the placement of peripheral water injectors maintains the reservoir pressure above minimum miscibility pressure. All reservoirs show complex variable slope compositional gradients along thick oil columns of hundreds of feet (Fig. 1). To match the fluid behavior and the variation of fluid properties with depth, the equation of state needs at least nine components. The rock quality mainly is controlled by diagenesis. Thirteen rock types were modeled. The permeability can change up to four log cycles for the same porosity. Most of the reservoirs are highly heterogeneous, with features such as high-permeability streaks and baffle zones. A wide range of capillary pressure curves is present; these mainly depend on permeability and lithology. Most development wells were completed with inflow control devices (ICDs) to control gas and water breakthroughs and optimize oil production. The combination of numerous ICDs and long slanted production intervals (i.e., thousands of feet) make wellbore-reservoir coupling critical for proper history matching and forecasting. The model grid size in the horizontal direction is 328 ft, which is considered optimal according to simulated sensitivities. The vertical layering is very fine in order to capture the reservoir heterogeneity; the cell thickness ranges from 1 to 1.5 ft. This results in a model with 3.5 million active cells, which makes the simulation performance and run time very challenging when coupled with compositional gradient and ICDs. In this section of the complete paper, the authors review why accurate phase labeling is important in compositional simulation and how it can lead to convergence problems, particularly for cases with gas injection. A traditional method is compared with an ML method on a simple model using a complex dummy fluid model to highlight issues that may arise in more-complex simulation models.