Integrating petrophysical data into efficient iterative cluster analysis for electrofacies identification in clastic reservoirs

Efficient iterative unsupervised machine learning involving probabilistic clustering analysis with the expectation-maximization (EM) clustering algorithm is applied to categorize reservoir facies by exploiting latent and observable well-log variables from a clastic reservoir in the Majnoon oilfield, southern Iraq. The observable well-log variables consist of conventional open-hole, well-log data and the computer-processed interpretation of gamma rays, bulk density, neutron porosity, compressional sonic, deep resistivity, shale volume, total porosity, and water saturation, from three wells located in the Nahr Umr reservoir. The latent variables include shale volume and water saturation. The EM algorithm efficiently characterizes electrofacies through iterative machine learning to identify the local maximum likelihood estimates (MLE) of the observable and latent variables in the studied dataset. The optimized EM model developed successfully predicts the core-derived facies classification in two of the studied wells. The EM model clusters the data into three distinctive reservoir electrofacies (F1, F2, and F3). F1 represents a gas-bearing electrofacies with low shale volume (V_sh) and water saturation (S_w) and high porosity and permeability values identifying it as an attractive reservoir target. The results of the EM model are validated using nuclear magnetic resonance (NMR) data from the third studied well for which no cores were recovered. The NMR results confirm the effectiveness and accuracy of the EM model in predicting electrofacies. The utilization of the EM algorithm for electrofacies classification/cluster analysis is innovative. Specifically, the clusters it establishes are less rigidly constrained than those derived from the more commonly used K-means clustering method. The EM methodology developed generates dependable electrofacies estimates in the studied reservoir intervals where core samples are not available. Therefore, once calibrated with core data in some wells, the model is suitable for application to other wells that lack core data.