Well-log facies classification using an active semi-supervised algorithm with pairwise constraints

Abstract

SUMMARY We present a technique for lithofacies classification of well-log data using an active semi-supervised algorithm. This method considers both the input of domain experts and the distribution characteristics of well-log properties. It aims to obtain lithofacies that are more geologically meaningful and seismically interpretable than the conventional clustering methods. We impose guidance from experts (e.g. geologist, petrophysicist and seismic interpreter) as pairwise constraints. The acquired constraints were incorporated into facies classification in two ways: modification of the objective function and optimization of the classification subspace. An iterative expectation-maximization (EM) algorithm was used to minimize the objective function. We applied the method to a set of well logs from the Glitne field, North Sea, where six lithofacies had been defined initially. Classification results illustrated that facies predicted with the semi-supervised approach achieved good matches with true labels. Comparisons among different methods (semi-supervised method, quadratic determinant analysis and expectation-maximization with Gaussian mixture model algorithm) also demonstrated that the proposed method significantly outperformed the others. We also tested a scenario with five facies, where we combined silty shale and shale into one group due to significant overlap in the elastic domain. Results demonstrated that the semi-supervised approach produced facies that were more consistent with expert intention, and they were more geologically interpretable. The techniques and results illustrated here could be performed in different types of reservoir facies classification, and the facies classified using semi-supervised algorithm honours the input of the users and data characteristics.