A Mixture-Of-Experts-Assisted Optimization Protocol for CO2-EOR and Storage Decision-Making

Abstract

CO2 enhanced oil recovery and geological storage require the simultaneous optimization of oil production, CO2 storage, and project economics, while the relative importance of these objectives and the preferred injection strategy can change throughout the project life cycle. Conventional multi-objective optimization generally determines a single Pareto optimal set using fixed design parameters and therefore cannot fully represent these time-dependent changes. This work develops a mixture-of-experts-assisted dynamic multi-objective optimization protocol for adaptive CO2-WAG design and decision-making. The mixture-of-experts model combines radial basis function, residual deep neural network, and Kolmogorov-Arnold network sub-experts through a lightweight gating network and serves as a surrogate for repeated high-fidelity reservoir simulations. Multi-objective particle swarm optimization is then performed at prescribed time steps to generate time-dependent Pareto optimal sets. An entropy weighting method evaluates the evolving importance of cumulative oil production, CO2 storage, and project net present value and ranks the non-dominated solutions to identify the preferred operating strategy. The workflow is first validated using a Xinjiang injector-producer well-pair model and is then applied to the UNISIM-IV benchmark reservoir over a 20-year CO2-WAG development period. Blind testing confirms that the mixture-of-experts model consistently outperforms the individual sub-experts while introducing only a small number of additional learnable parameters. For UNISIM-IV, the model achieves an R2 of 0.9932 and an MRE of 1.34% for cumulative oil production, and an R2 of 0.9902 and an MRE of 6.83% for CO2 storage. The dynamic optimization reveals the temporal evolution of the Pareto optimal set, objective weights, and CO2-WAG design parameters, allowing the operating strategy to adapt as reservoir performance changes. Compared with the operator-designed base case, the dynamic protocol increases cumulative oil production, CO2 storage, and project NPV by approximately 67%, 14%, and 78%, respectively. More importantly, compared with conventional static optimization using a single set of control parameters, the proposed protocol increases the final project NPV by approximately 9.4% while maintaining nearly the same cumulative oil production, although the final CO2 storage is reduced by approximately 12%. These results demonstrate that dynamically adjusting the injection strategy provides a more favorable economic trade-off than a fixed optimized design and provides a quantitative basis for determining when a CO2-EOR project should progressively transition toward storage-oriented operation.