Abstract
Water flooding remains one of the most widely used improved oil recovery (IOR) methods, even after decades of hydrocarbon production and the application of various IOR techniques. However, real reservoir simulations are computationally expensive and time-consuming due to the complexity of reservoir models. To address this, researchers increasingly rely on artificial intelligence and neural network-based proxy models as faster alternatives. While proxy models reduce computational effort, their ability to predict reservoir behavior remains a challenge reliably. This study evaluates the reliability of proxy models in optimization tasks. The Punq-S3 benchmark reservoir model was used to optimize well control parameters and maximize net present value (NPV) through two approaches: one based on proxy models and the other on full-physics reservoir simulation. Four deep learning algorithms (ANN, LSTM, GRU, and EL) were combined with two design of experiment techniques (Taguchi and Latin hypercube sampling) to generate eight proxy models. The particle swarm optimization and Bayesian optimization algorithms were selected to optimize the injection and production strategy for the two approaches. after optimization and comparison of the results, it was observed that despite the constructed proxy models exhibited acceptable predictive performance in terms of statistical metrics (e.g., RMSE, R²), the optimization results they yielded significantly deviated from those based on the full-physics reservoir model. This reveals a critical trade-off between computational efficiency and decision-making reliability. While proxy models offer a cost-effective alternative for rapid predictions, caution is needed when applying them to optimization tasks, where minor prediction errors can lead to substantially different outcomes. This study highlights the importance of evaluating proxy models not only based on accuracy metrics but also on their ability to capture optimization-relevant dynamics.