Comparison of Clustering Techniques to Define Chemofacies: Case Study for Mississippian Rocks in the STACK Play, Oklahoma
From reducing uncertainty in well correlations to identifying target zones with low and high reservoir quality, chemostratigraphy has demonstrated to be an excellent tool in geosciences. Chemofacies, an analog of lithofacies, are characterized by a signature composition of 30 elements obtained with X-Ray fluorescence (XRF) spectroscopy. However, a standard definition of the chemofacies is ambiguous due to the different techniques available for clustering analysis.We aim to create a methodology for clustering Mississippian strata with a similar elemental composition in the Anadarko basin. This result in the chemofacies that are used for: well correlations, paleoenvironment interpretations, identify landing zones and refine a sequence stratigraphic framework. To address the issue of chemofacies clustering ambiguity, we used different unsupervised learning techniques in over 1000 analyses of XRF spectroscopy, acquired for Mississippian strata in 4 cores located in the STACK play, Oklahoma. We lead with different questions to define a methodology for defining the chemofacies. The first issue we deal with is the selection of the elements to be clustered. Sometimes chemofacies are based only on the elements that have been used in the literature as a proxy for any geological parameter. For example, Sr, Mg and Ca, as carbonates proxies. But, with the objective to incorporate information that might escape traditional geological inference, we also used principal component analysis (PCA) as a preprocessing step before clustering the elements. The next challenge we address is to analyze which clustering algorithm can better represent Mississippian strata. We compare the results of Hierarchical cluster analysis (HCA), K-means, Self-organizing map (SOM), and Density-based spatial clustering (DBSCAN). Then, we used PCA to geologically constrain the clusters and define the chemofacies. Finally, the chemofacies were compared with thin sections and well logs. The analysis we performed allowed us to define the most appropriate workflow that honors the geology embedded in the lithofacies. The selection of unsupervised learning algorithm is based both in the resulted chemofacies and the clustering objectives. We propose that the segmentation of massive gravity flows facies from hemipelagic facies can be achieved with two clusters. However, more clusters are necessary if the objective is to identify lower and higher reservoir quality intervals within these two main clusters.
AAPG Datapages/Search and Discovery Article #90350 © 2019 AAPG Annual Convention and Exhibition, San Antonio, Texas, May 19-22, 2019