Facies classification is a crucial task which can improve the chances of success of a well significantly. The relevant classification algorithms take well logs as inputs and classify the formation into distinctive clusters or electrofacies. Integrating the electrofacies with core measurements can lead to an understanding of the geological facies. We develop a general-purpose workflow for unsupervised electrofacies classification, which takes well logs as inputs and can be used for different application scenarios. The clustering is performed using the Gaussian mixture model approach. The optimal number of clusters is automatically determined ensuring repeatable clustering results from multiple realizations of the classification workflow. The workflow was applied on field data from off-shore Norway. We observe high similarity in the resulting facies with the ones determined visually by the field geologist from core data, by comparing their permeability-porosity relationships. This new approach removes the user intervention in the workflow and provides a robust solution for automating the electrofacies classification processing.


Facies classification is a key element in the evaluation of petrophysical formations and in reservoir characterization. Electrofacies are defined as clusters of similar log responses in a well or a set of wells and their combination with core measurements can lead to geological facies, which can represent series of petrophysical properties. There has been significant progress towards developing automated workflows for facies classification (Busch et al, 1987; Lim et al. 1997; Rabaute, 1998; Qi and Carr, 2005; Skalinski et al., 2006; Tang et al., 2011).

There are three main challenges in electrofacies classification. First, the fact that most of the times there are no labeled data necessitates to use an unsupervised classification method. There are various unsupervised learning algorithms like the k-means (Lloyd, 1982) or the hierarchical clustering algorithm (Ward, 1963) to perform classification. However, these algorithms perform "hard" assignment of data points to clusters, in which each data point is associated uniquely with one cluster (Bishop, 2006) and they do not consider the fact that field data can have some uncertainty over the clusters they are assigned. Second, the optimal number of clusters is usually unknown and thus is required to be an input given by the user. Various approaches have been developed to avoid the user’s subjectivity in the choice of the optimal number of clusters and automate the process. Some of the most common ones are the Bayesian Information Criterion (Schwarz, 1978) and the Cross-Entropy Clustering (Tabor and Spurek, 2014), which however do not provide a universally robust solution. Finally, it is very common different realizations of the classification algorithm to give different clustering results, even if the input logs and the algorithm parameters are kept the same for all realizations. This is because each of the input parameters of the algorithm is initialized randomly for each realization and as a result the algorithm converges to a different value of loglikelihood and the clustering is different each time consequently.

This content is only available via PDF.
You can access this article if you purchase or spend a download.