Cell-level deep learning as proxy model for reservoir simulation and production forecasting

Optimizing strategies in the Oil and Gas Industry, particularly within reservoir engineering and management, remains a significant challenge due to the prohibitive computational time costs and high resource demands of current simulation methods, even for medium-sized reservoirs. Notably, existing scientific approaches have not leveraged Deep Neural Networks (DNNs) for fine-grained predictions at the grid-cell level. This thesis introduces a novel approach that integrates DNNs with a Design of Experiments framework to develop a proxy model for reservoir simulation software. The methodology includes a robust feature selection process, model design, and training strategy, supplemented by comprehensive statistical evaluations and graphical tools. The proposed proxy models are validated using four distinct industrial scenarios based on the SPE9 black oil benchmark, incorporating production and injection wells across diverse temporal samples. The results demonstrate a significant improvement in computational efficiency without compromising accuracy, achieving over 80% accuracy across all scenarios, and reaching up to 99.9% under specific conditions. These findings highlight the potential of DNN-based proxy models to transform reservoir management practices, offering a scalable and resource-efficient alternative to traditional numerical simulations.