Advanced Static Model Screening Using Non-Supervised Machine Learning Techniques

Effective selection of static reservoir models is a critical step in enhancing the accuracy, efficiency, and reliability of reservoir simulation workflows. Traditional approaches often rely on volumetric properties alone, which may not fully capture the dynamic behaviour of the reservoir under production. This study introduces a novel methodology that integrates streamline simulation outputs into the model selection process, offering a more performance-oriented perspective. By applying Principal Component Analysis (PCA) to streamline-derived metrics—such as sweep efficiency, flow connectivity, and drainage patterns—the method reduces the dimensionality of the dataset while preserving the most informative features. Hierarchical clustering is then used to group realizations based on their dynamic similarity, enabling a structured and data-driven classification of the model ensemble. A key innovation of this approach lies in its multi-scale integration of field-level and well-level data, which allows for a more comprehensive understanding of reservoir response. Field-level metrics provide insight into global reservoir performance, while well-level indicators capture localized dynamics such as individual well contributions, connate water production, and breakthrough timing. By combining these two scales of analysis, the methodology ensures that selected models reflect both the broader reservoir trends and the nuanced output pattern of specific wells. This dual-layered perspective enhances the fidelity of reservoir characterization and supports more informed decision-making, particularly in complex or heterogeneous reservoirs where local variations can significantly impact overall performance. To support interpretation and streamline the selection process, an interactive visualization dashboard is employed. This tool enables real-time exploration of the clustered ensemble, with each realization color-coded by its assigned cluster label. Users can dynamically filter, compare, and validate models, facilitating the generation of a shortlist of representative cases for full-physics simulation. By focusing on a reduced set of high-impact realizations, the workflow significantly lowers computational costs and accelerates simulation timelines—without compromising the diversity or representativeness of the model set. Ultimately, this integrated methodology offers a powerful and scalable solution for optimizing reservoir development strategies, improving operational planning, and enabling more agile, data-driven reservoir management.