Multi-Objective Optimization of CO₂ Injection Process into Oil Reservoirs Using Machine Learning Algorithms: Incorporating Carbon Sequestration Mechanisms

Capture and storage of CO2 in underground geological formations has been identified as a sustainable solution for mitigating the effects of greenhouse gases. Combining this CO2 sequestration with enhanced oil recovery (EOR) processes can reduce the economic risk of carbon capture and storage (CCS). Injecting CO2 alternately with water (water alternating gas or WAG) is recognized as one of the most effective methods for increasing oil production and enhancing CO2 sequestration. This study aims to optimize the CO2 injection process into oil reservoirs using the WAG method, explicitly focusing on incorporating various carbon sequestration mechanisms. Due to the inherent complexities of the WAG injection process and the conflicts of interest between specific CO2 sequestration mechanisms and cumulative oil production (COP), there is a need for a practical multiobjective optimization approach. In this study, based on the mechanisms of CO2 trapping in the oil reservoir, three different objective functions representing the moles of CO2 trapped in different phases within the reservoir, along with the COP objective function, were considered. Using reservoir simulation, 366 realizations were designed based on seven decision variables, and the four mentioned objective functions were calculated. Initial correlation analysis among the objective functions confirmed a conflict of interest between the COP objective function, the CO2 trapped in oil (CTO) and water (CTW) phases, and conflicts between the trapping mechanisms. Multiple proxy models were trained using the created data set and two machine learning methods, XGBOOST, and neural networks. Ultimately, a neural network with an R2 of 0.9886 for the training phase and 0.9562 for the test phase was selected as the validated proxy model. Optimizing solutions were evaluated by integrating the proxy model with three multiobjective optimization algorithms (NSGA-II, PESA-II, and MOPSO). Due to the conflict of interest among the objective functions, optimization was conducted using two different cost function settings, ensuring that all potential optimal solutions were identified. The results demonstrated that the shape of the Pareto front and the arrangement of the optimal solutions change when CO2 trapping mechanisms are applied, compared to previous optimization approaches. The CO2 sequestration objective function is significantly better optimized when these trapping mechanisms are included in the optimization process. Therefore, incorporating various CO2 trapping mechanisms into the CO2–WAG process optimization framework is essential to avoid overlooking potential solutions.