Hybrid Multiscale Simulation Enhanced by Super – Resolution for Multiphase Flow in High – Contrast Heterogeneous Reservoirs

Abstract Robust field development strategies rely on optimizing well placement and control through extensive high-resolution reservoir simulations. However, the computational cost of such simulations can be prohibitively high. To address this challenge, this study develops an efficient dynamic well-control optimization workflow by leveraging a super-resolution deep neural network (SR-DNN) model to alleviate the computational burden. The SR-DNN is trained on paired coarse- and fine-scale reservoir simulations to learn the nonlinear mapping between coarse-scale state variables (pressure and saturation) and their fine-scale counterparts. The SR-DNN is integrated with an efficient coarse-scale simulator, forming a hybrid forward model that accurately predicts production well rates. The workflow is validated through dynamic well-control optimization for a water-flooding process in a channelized reservoir. The SR-DNN achieves excellent prediction accuracy, with average root mean square error (RMSE) of 76.1 psi for pressure and 0.046 for saturation throughout 10 years of prediction horizon. The structural similarity index measure (SSIM) further confirms high fidelity, averaging 0.9899 for pressure and 0.9129 for saturation. The hybrid forward model demonstrates strong agreement with fine-scale simulations, yielding L2-norm relative error of 1.79% and 5.54% for oil and water production cumulative, respectively. Additionally, the hybrid model achieves an NPV with an R2 score of 0.92 relative to the fine-scale simulation benchmark, while delivering a 16× speedup with an upscaling factor of 5 × 5. This proposed work features a novel integration of deep-learning-based super-resolution with coarse-scale reservoir simulation which can be readily used for fine-scale dynamic well-control optimization. Unlike purely data-driven approaches, the method preserves physical consistency by leveraging coarse-scale simulations for guidance. The proposed framework advances reservoir simulation by combining computational efficiency with high accuracy, enabling large-scale well-control optimization with practical feasibility.