The genetic clustering of oil samples, as well as of petroleum source rocks are vital tasks for hydrocarbon exploration and other related fields. This practice is commonly named oil-oil or oil-source correlation and it is based on the identification of compositional similarities. It reveals sample affinities regarding their origin, thermal maturity stage, biodegradation stage, family affiliation etc. The results from such studies benefit tasks like the definition of petroleum systems, the identification of source rock paleo-depositional environments across study areas and the affinity of oil shows to source rocks. So far, methods such as principal component analysis (PCA), hierarchical clustering (HCA), k-means clustering, and multidimensional scaling (MDS) have been extensively used in chemometrics for such studies, including biomarkers and normal alkanes of the saturated and aromatic fractions i.e.,. Although widely used, drawbacks do exist in these methods, basically related to the methods used for distance measurements, the lack of compositional information (biodegraded samples), or the incorrect use of biomarker ratios. Contributing to this line of research, we present a new method for such clustering studies; a direct multiway spectral clustering that allows the automatic selection of the optimal number of clusters. Spectral clustering is a long-established graph-based method that enjoys reasonable computation time, implementation simplicity, and overcomes the NP-hardness of other graph-theoretic algorithms. Nonlinear reformulations of the method have proven to lead to improved clustering assignments and normalized variants have been shown to lead to the modularity maximization of the resulting clusters. These allow the automatic detection of the optimal number of groups present in the graph.
Please find the full abstract below.