A Surrogate-Assisted Framework for Multiobjective Production Optimization of Graded-Viscosity Polymer Flooding

Summary Due to its superior capacity in improving swept volume and enhancing oil displacement efficiency, polymer flooding has proved to be a promising alternative to further increase oil recovery in high-water-cut mature oil fields. However, the traditional optimization framework, which strongly depends on expert knowledge and high-fidelity simulators, is computationally intensive and inefficient in addressing complex engineering constraints. Balancing efforts to decrease water cut and increase oil production while reducing polymer consumption poses significant challenges, which greatly hinder the optimal design of polymer flooding injection schemes in practice. To address this problem, we propose a novel surrogate-assisted framework for multiobjective production optimization of graded-viscosity polymer flooding in high-water-cut mature reservoirs by incorporating the graded-viscosity constraint as an explicit feasible region throughout the optimization search. In this study, we first develop an efficient methodology that couples reservoir numerical simulation for graded-viscosity polymer flooding with a surrogate model and then iteratively compute the Pareto front of a multiobjective optimization model by developing a graded-viscosity constrained classifier-assisted rank-based learning and local model-based multiobjective evolutionary algorithm (GC-CLMEA). The commonly used technique for order preference by similarity to ideal solution (TOPSIS) is ultimately used to obtain the optimal injection schemes for a synthetic reservoir model. The results show that, as the polymer concentration gradient in the main slug increases from 0 mg/L to 1,400 mg/L, the final recovery factor generally increases from 53.06% to 64.17%. Notably, the recovery increment gradually decreases in the higher-gradient range. Meanwhile, a medium concentration gradient of 600–800 mg/L is more conducive to the mobilization of low-permeability layers. These results indicate that a moderate increase in concentration gradient helps improve sweep efficiency, whereas an excessively high gradient weakens the ability of the displacing fluid to advance into the deeper parts of low-permeability layers. The proposed GC-CLMEA demonstrates faster convergence and better overall performance than the comparison algorithms, reaching the same hypervolume (HV) in 261 iterations, which is 581 fewer iterations than CLMEA. Compared with the basic scheme, the optimal injection scheme increases cumulative oil production from 3.3203×105 m3 to 3.4849×105 m3, improving the recovery factor by 2.95%, and reduces polymer consumption from 16.8831×105 kg to 15.8727×105 kg. Further application to a real reservoir block confirms that GC-CLMEA can maintain stable HV evolution and generate a high-quality Pareto solution set under complex reservoir conditions.