Unsupervised identification of electrofacies employing machine learning

Machine learning techniques are widely used in petrophysicics and geophysics to solve complex and non-linear problems of practical importance. In particular, numerous applications for identifying electrofacies from well logs have been conducted. However, there is no unique approach for reliable automatic classification of electrofacies as the accuracy of the applied algorithms may vary depending on data and initial conditions. To overcome instability in outcomes from various algorithms, we suggest applying different clustering techniques to log data, in a way similar to the popular method of supervised classification ensemble learning. Such an ensemble of clustering outputs are further integrated into unique classes (electrofacies) for subsequent automated identification of lithofacies. Here we apply three different clustering algorithms, namely, Spectral Clustering Self-Organizing Map and k-means, in order to reliably classify electrofacies at a petroleum exploration well Lauda-1 drilled in the Northern Carnarvon Basin (Western Australia). The clustering outputs integrated into electrofacies are validated using an expert facies classification. We show that some facies identified by the expert are not distinguished as separate classes, at least for the chosen well and selected logs. The established electrofacies can further be assigned to conventional lithofacies. This requires creating an expert system which is currently under development.