Wireline logs have been utilized to indirectly estimate various reservoir properties, such as porosity, permeability, saturation, cementation factor, and lithology. Attempts have been made to correlate Gamma-ray, density, neutron, spontaneous potential, and resistivity logs with lithology. The current approach to estimate grain size, the traditional core description, is time-consuming, labor-intensive, qualitative, and subjective. An alternative approach is essential given the utility of grain size in petrophysical characterization and identification of depositional environments.
This paper proposes to fill the gap by studying the linear and nonlinear influences of wireline logs on reservoir rock grain size. We used the observed influences to develop and optimize respective linear and machine learning models to estimate reservoir rock grain size for a new well or targeted reservoir sections. The linear models comprised logistic regression and linear discriminant analysis while the machine learning method is random forest (RF). We will present the preliminary results comparing the linear and machine learning methods. We used anonymized wireline and archival core description datasets from nine wells in a clastic reservoir. Seven wells were used to train the models and the remaining two to test their classification performance. The grain size-types range from clay to granules. While sedimentologists have used gamma-ray logs to guide grain size qualification, the RF model recommended sonic, neutron, and density logs as having the most significant grain size in the nonlinear domain.
The comparative results of the models' performance comparison showed that considering the subjectivity and bias associated with the visual core description approach, the RF model gave up to an 89% correct classification rate. This suggested looking beyond the linear influences of the wireline logs on reservoir rock grain size. The apparent relative stability of the RF model compared to the linear ones also confirms the feasibility of the machine learning approach.
This is an acceptable and promising result. Future research will focus on conducting more rigorous quality checks on the grain size data, possibly introduce more heterogeneity, and explore more advanced algorithms. This will help to address the uncertainty in the grain size data more effectively and improve the models performance. The outcome of this study will reduce the limitations in the traditional core description and may eventually reduce the need for extensive core description processes.