Lithological facies classification using well logs is essential in the reservoir characterization. The facies are manually classified from characteristic log responses derived, which is challenging and time consuming for geologically complex reservoirs due to high variation of log responses for each facies. To overcome such a challenge, machine learning (ML) is helpful to determine characteristic log responses. In this study, we classified the lithofacies by applying ML to the conventional well logs for the volcanic formation, onshore, northeast Japan. The volcanic formation of the Yurihara oil field is petrologically classified into five lithofacies: mudstone, hyaloclastite, pillow lava, sheet lava, and dolerite, with pillow lava being predominant reservoir. The former four lithofacies are the members of the volcanic system in Miocene, and dolerite randomly intruded later into those. Understanding the distribution of omnidirectional tight dykes at the well location is important for the estimation of potential near-lateral seal distribution compartmentalizing the reservoir. The facies are best classified by core data, which are unfortunately available in a limited number of wells. The conventional logs, with the help of the borehole image log, have been used for the facies classification in most of the wells. However, distinguishing dolerite from sheet lava by manual classification is very ambiguous, as they appear similar in these logs. Therefore, automated clustering of well logs with ML was attempted for the facies classification. All the available log data was audited in the target well prior to applying ML. A total of 10 well logs are available in the reservoir depth interval. To prioritize the logs for the clustering, the information of each log was first analyzed by Principal Component Analysis (PCA). The dimension of variable space was reduced from 10 to 5 using PCA. Final set of 5 variables, gamma-ray, density, formation photoelectric factor, neutron porosity, and laterolog resistivity, were used for the next clustering process. ML was applied to the selected 5 logs for automated clustering. Cross-Entropy Clustering (CEC) was first initialized using k-means++ algorithm. Multiple initialization processes were randomly conducted to find the global minimum of cost function, which automatically derived the optimized number of classes. The resulting classes were further refined by the Gaussian Mixture Model (GMM) and subsequently by the Hidden Markov Model (HMM), which takes the serial dependency of the classes between successive depths into account. Resulting 14 classes were manually merged into 5 classes referring to the lithofacies defined by the borehole image log analysis. The difference of the log responses between basaltic sheet lava and dolerite was too subtle to be captured with confidence by the conventional manual workflow, while the ML technique could successfully capture it. The result was verified by the petrological analyses on sidewall cores (SWCs) and cuttings. In this study, the automated clustering with the combination of several ML algorithms was demonstrated more efficient and reasonable facies classification. The unsupervised learning approach would provide supportive information to reveal the regional facies distribution when it is applied in the other wells, and to comprehend the dynamic behavior of the fluids in the reservoir.
A Practical Method for the Automatic Recognition of Rock Structures in Panoramic Borehole Image during Deep-Hole Drilling Engineering