Indirect Estimation of Clastic Reservoir Rock Grain Size from Wireline Logs Using a Supervised Nearest Neighbor Algorithm: Preliminary Results

Abstract

Reservoir rock textural properties such as grain size are typically estimated by direct visual observation of the physical texture of core samples. Grain size is one of the important inputs to petrophysical characterization, sedimentological facies classification, identification of depositional environments, and saturation models. A continuous log of grain size distribution over targeted reservoir sections is usually required for these applications. Core descriptions are typically not available over an entire targeted reservoir section. Physical core data may also be damaged during retrieval or due to plugging. Alternative methods proposed in literature are not sustainable due to their limitations in terms of input data requirements and inflexibility to apply them in environments with different geological settings. This paper presents the preliminary results of our investigation of a new methodology based on machine learning technology to complement and enhance the traditional core description and the alternative methods. We developed and optimized supervised machine learning models comprising K-nearest neighbor (KNN), support vector machines (SVM), and decision tree (DT) to indirectly estimate reservoir rock grain size for a new well or targeted reservoir sections from historical wireline logs and archival core descriptions. We used anonymized datasets consisting of nine wells from a clastic reservoir. Seven of the wells were used to train and optimize the models while the remaining two were reserved for validation. The grain size types range from clay to pebbles. The performance of the models confirmed the feasibility of this approach. The KNN, SVM, and DT models demonstrated the capability to estimate the grain size for the test wells by matching actual data with a minimum of 60% and close to 80% accuracy. This is an accomplishment taking into account the uncertainties inherent in the core analysis data. Further analysis of the results showed that the KNN model is the most accurate in performance compared to the other models. For future studies, we will explore more advanced classification algorithms and implement new class labeling strategies to improve the accuracy of this methodology. The attainment of this objective will further help to handle the complexity in the grain size estimation challenge and reduce the current turnaround time for core description.