Application of Machine Learning Algorithms in Hydrocarbon Exploration and Reservoir Characterization

Abstract

This dissertation presents novel approaches to evaluate complex seismic and well-log data using machine learning algorithms with examples from two different hydrocarbon fields. The applicability of these algorithms for predicting and classifying direct or indirect hydrocarbon indicators are assessed and compared to knowledge-driven methods. The efficacy of the various techniques leads to recommendations for utilizing machine learning algorithms in well planning or later cycle hydrocarbon-field development. In the first study in this dissertation, application of a model-based artificial neural network is compared to the performance of a prestack simultaneous inversion method in predicting hydrocarbon presence in the Heidrun Field, offshore Norway. Low-frequency initial models were used to create 3D Poisson’s ratio models to reflect the fluid within this field and the results were compared based on the accuracy and generalization power of the two methods. The results of both methods confirmed Poisson’s ratio to be a good direct hydrocarbon indicator within the wells used from this field. The direct dependency of the inversion method on the provided input constraints, however, can raise the risk for well planning decisions beyond the known zones. The generalize regression neural network results better matched the observations at the training wells and provided a lower risk of false discoveries in delineating favorable zones beyond the drilled wells. The second study was conducted with the aim of classifying different facies from well logs in wells of the Heidrun Field and in the Kupe Field, offshore New Zealand. Different machine learning approaches were utilized in this study and to investigate quantitatively and qualitatively the accuracy and stability of their predictions. Both supervised methods could successfully predict hydrocarbon-bearing units, with the bagged tree algorithm having a higher overall, and hydrocarbon-related, accuracy rate. Application of the bagged-tree algorithm showed a very low false discovery rate for oil sands and no false discoveries for gas sands in the Heidrun Field. A misclassification of oil sands as brine sands in one Heidrun well is in agreement with relatively high Poisson’s ratios as discussed in the first study. Qualitative investigations of Kupe Field results also demonstrated accurate prediction of hydrocarbon-bearing units, including a shaly hydrocarbon sand class defined for low-quality reservoir sands. Hydrocarbon shows reported in one well that were not predicted by the algorithm, in fact, occur in a very low-porosity section of the reservoir that was not identified as reservoir in reports either. In the last study, the classifications of the litho-fluid facies were extended to three dimensional models using two machine learning methods and were compared with a knowledge-driven approach. The results were examined through a probabilistic approach to reflect the uncertainty of the predicted classes. The probabilistic neural network and the bagged-tree algorithm successfully predicted the variations of litho-fluid facies, especially for hydrocarbon units. Both methods predicted gas sands in certain parts of the field, away from control points, with similar form and lateral dimension. By comparing the results in predicting oil sands and shale, we interpret the bagged-tree method to be more adherent to the known parameters set by the interpreter, such as the OWC and the target classes. Predictions from the probabilistic neural network, however, can deviate from the target facies even close to the wells on which it has been trained. The efficiency of machine learning techniques in increasing the prediction accuracy and decreasing the procedure time, and their objective approach toward the data, make it highly desirable to incorporate them in seismic data analyses. Along with the emphasis on the application of machine learning techniques in the study of subsurface properties, this dissertation presents frameworks for utilizing these techniques as new tools for the interpreter, not as a replacement. The knowledge of the data analyst about the field, and the selection and preparation of the attributes and application of the appropriate algorithm are all crucial factors in this procedure.