Improving Electrofacies Prediction by Combining Supervised and Unsupervised Learning Methods

Electrofacies classification from well logs is an indispensable part of seismic interpretation and is important in the determination of sequence stratigraphy, and ultimately reservoir characterization. Although there have been improvements in the tools used to perform this task, it remains laborious, subjective, and error-prone. Achieving a proper classification is complicated by increasing dataset sizes as well as the need for correlated multidisciplinary models. Recent developments in machine learning provide an opportunity to assist interpreters in accomplishing this task while also improving the accuracy of classification results. Applications of machine learning methods for automating facies classification from well logs have previously been explored, however these have mostly focused on evaluations or comparisons of individual algorithms or of ensembles of homogeneous agents. The proposed methods combine heterogeneous agents to enhance prediction accuracy and expedite the assessment of large datasets. This approach seamlessly integrates supervised and unsupervised learning techniques, effectively capitalizing on their individual strengths and mitigating their inherent limitations. The overarching objective is to offer valuable support to geoscientists by not only improving prediction accuracy beyond the constraints of current methodologies, but also by significantly accelerating the evaluation process for progressively expanding datasets. To accomplish this, supervised learning, which establishes a direct mapping from the data domain to the solution domain while introducing some bias to generalize the mapping, is coupled with unsupervised learning, which operates without reliance on similar generalization bias or predefined training data but does not offer a direct mapping between the data and solution domains. This fusion is achieved through the utilization of a joint probability density function (PDF) derived from the supervised classification. The PDF serves to guide the identification of clusters outlined by unsupervised learning. This multi-agent approach can effectively detect bias introduced during training for as many as one in five samples present in well log data, and forms the basis for generating a probability distribution for individual samples, rather than simply assigning a discrete classification for each sample. Consequently, this distribution proves valuable in more accurately representing the continuous nature of well log signals, and captures the intrinsic continuity present within lithological regimes. A modified version of this approach can be applied to rapidly identify regions of interest within complex depositional environments which may involve hundreds or even thousands of boreholes equipped with extensive suites of well log data. This innovative adaptation streamlines the evaluation process, reducing an interpretation task that could have consumed months to just hours. As a result, geoscientists can dramatically scale up their interpretation efforts, increasing their output substantially. The research concludes by establishing a foundation for comparing the predictive accuracy of the proposed approach with that of traditional petrophysical analysis and of core interpretation; where these proven methods serve as benchmark references for the evaluation.