Rapid evaluation and optimization of carbon dioxide-enhanced oil recovery using reduced-physics proxy models

The carbon dioxide (CO2) injection in oil reservoirs is a promising enhanced oil recovery method to reduce carbon emissions into the atmosphere. The CO2 injection simulation and optimization require a large computation time, especially in real large-scale oil reservoirs. This paper integrates experimental design and machine learning to construct a reduced-physics surrogate model alternative to the complex reservoir simulator. This surrogate model was used to evaluate and optimize the CO2-assisted gravity drainage (CO2-GAGD) process, applied to a clastic reservoir in the Rumaila oil field. In the GAGD process optimization, five operational decision parameters controlling the production and injection activities were manipulated to attain optimal future oil production. Hundreds of simulation runs were created by Latin Hypercube Sampling to build a proxy-based optimization. The optimal scenario increased the cumulative oil production by 416 MMSTB at the end of 10-year prediction period. Finally, four machine learning (ML) algorithms were built as proxy surrogates alternative to the complex compositional reservoir simulations: Quadratic Equation (QM), FUzzy logic-GEnetic algorithm (FUzzy-GEnetic), Multivariate Additive Regression Splines (MARS), and Generalized Boosted Modeling (GBM). To show how robust using the ML approaches for the fast CO2-GAGD process optimization, the random subsampling cross-validation was adopted to conclude the optimum proxy metamodel that provides the lowest mismatch between the proxy- and simulator-based cumulative oil production. The GBM algorithm achieved the highest adjusted-R-2 (0.9973) and lowest root mean square prediction error ( 3.704 x 1 0 6 ) $(3.704\times 1{0}<^>{6})$ to produce the most accurate metamodel. The resulting accurate GBM proxy metamodel can be used for fast optimization of the CO2-GAGD process. Specifically, GBM-metamodel should lead to achieving higher optimal reservoir flow responses, especially with running a large number of simulation runs. Consequently, this proposed trained GBM-ML proxy metamodel could be used for gas injection optimization and uncertainty assessment with far-less computational effort than with the conventional approaches, which use a complex flow simulator.

Automated summary

This paper proposes a simplified physics surrogate model constructed using experimental design and machine learning, which is used to evaluate and optimize the CO2-assisted gravity drainage process. By manipulating operational decision parameters, the goal of increasing oil production in the reservoir is achieved. Additionally, four machine learning algorithms were built as proxy surrogates, and the most accurate metamodel was selected through cross-validation.