Machine learning clustering methods offer the potential for recognition and separation of facies based on core or well-log data. This is a particular problem for carbonate rocks because diagenesis produces a wide range of rock microstructures and transport properties. In this work we use a large database of high quality poroperm, electrical, mercury injection capillary pressure and nuclear magnetic resonance spectroscopy measurements (307 core samples), representing 5 stratigraphically defined facies, as well as well log data to examine facies-recognition abilities using 8 different machine learning clustering approaches and a redundancy of 10 to ensure statistically valid results, resulting in a total of over 990 clustering runs. For a 3 cluster problem, we find that the Expectation Maximisation (92.57% success) and two types of Kmeans approaches (89.60% and 91.09%) provide the best methods. Further testing using the best of these shows that the quality of the input parameter (attribute) matters more than the number of attributes used, with the power of attributes in decreasing order being porosity, cementation exponent, permeability, pore throat diameter and free fluid index, implying that some attributes can degrade clustering performance. Further tests show that there should be at least as many attributes as clusters, in which case the machine learning can be left to choose the final number of clusters, providing the best performance in this work (69.35% success for a five cluster problem), otherwise it is best to constrain the cluster number by supervision. Application of the results from the previous testing to a mixed carbonate tight carbonate well from the Butmah formation shows satisfactory determination of 4 petrofacies by clustering (up to 91.65%) when compared to petrofacies determined manually. However, the greater challenge of clustering 9 reservoir quality classes defined using a ternary petrofacies approach did not provide a successful result (<38% success rate).
CORE (COnnecting REpositories)
CORE (COnnecting REpositories)