Abstract.
This project explores the application of machine learning for petrophysical property characterization, focusing on pressure clusters within clastic sandstone reservoirs in the Niger Delta region. Pressure clusters, representing zones with similar pressure regimes and flow dynamics, serve as a powerful criterion for characterizing reservoir properties. By leveraging pressure clusters, this study captures spatial variability in reservoir properties more effectively, bypassing the need for labor-intensive sampling and core acquisition. A comprehensive dataset comprising petrophysical data from well logs and pressure measurements from Bottom-Hole Pressure (BHP) and Temperature tests was used. Machine learning algorithms, including Random Forest (RF) and Gaussian Mixture Models (GMM), were employed to identify and classify distinct pressure-based zones according to their petrophysical attributes. The results demonstrated exceptional accuracy, with a 99% prediction rate for classifying petrophysical properties across the reservoir. To ensure reliability, the machine learning results were cross-validated using traditional methods. The Reservoir Quality Index (RQI) and Flow Zone Indicator (FZI) techniques proposed by AlAjmi and Holditch (2000) were employed to independently assess the delineation of hydrocarbon flow units (HFUs). Remarkably, both approaches converged on the identification of two primary HFUs, confirming the consistency of the machine learning approach with conventional reservoir characterization techniques. Furthermore, RQI and FZI log-log plots showed a high degree of correlation (R² = 96%), underscoring the effectiveness of both methods in defining homogeneous zones. This project demonstrates the synergy between machine learning and traditional methods, emphasizing that machine learning offers a more efficient route to achieve the same outcomes. By employing pressure clusters as a criterion for petrophysical property delineation, machine learning enables rapid identification of distinct reservoir zones, even across unlogged intervals. Such capabilities provide a practical alternative to extensive coring, significantly reducing time and costs. Machine learning’s pattern recognition capabilities, particularly when applied to heterogeneous sandstone reservoirs, were pivotal in identifying zones with homogeneous flow and storage properties. Meanwhile, traditional methods like RQI and FZI remain invaluable for validating new approaches and ensuring robust conclusions. Together, these methodologies represent complementary tools in reservoir modeling, enabling precise rock typing and strategic reservoir management. This study highlights machine learning as a transformative tool that integrates seamlessly with established practices, bridging the gap between innovation and reliability in reservoir characterization
