A hybrid AI-genetic algorithm framework for the optimization of polymer flooding strategies: a numerical simulation-based approach.

Facing declining conventional resources, the oil industry requires advanced methods to maximize recovery. Polymer flooding is a key technique, but its optimization is hindered by complex parameter interactions and the high computational cost of traditional simulation. This study presents a novel solution: a hybrid AI-Genetic Algorithm (GA) framework that integrates numerical simulation with machine learning for efficient optimization. A large dataset of 960 core-scale simulation cases was generated to analyze key parameters like permeability and polymer concentration. The core innovation was the development of two neural networks, a Feedforward Neural Network (FNN) and an Elman Recurrent Neural Network (E-RNN), to act as fast proxy models. The E-RNN proved superior for forecasting dynamic production data, achieving exceptional accuracy (R² > 0.99) by effectively capturing time-dependent behaviors. This high-fidelity E-RNN proxy was then coupled with a GA for multi-objective optimization. Results showed that maximum oil recovery is achieved by maximizing permeability, injection rate, and polymer concentration while minimizing reservoir heterogeneity. Crucially, economic optimization revealed a different strategy, favoring a short, intensive injection period to maximize profit, highlighting a key technical-economic trade-off. The study successfully validated the framework's generalization capability. This work provides a powerful tool for accelerating polymer flooding design, with future efforts aimed at integrating laboratory data for calibration and scaling the application to full-field models.