Identifying Litho-Fluid Facies Using Machine Learning: A Case Study of Niger Delta Sandstone Reservoirs

Abstract To minimize uncertainties in reservoir characterization and modeling workflows, an accurate definition of reservoir heterogeneity, including lithologic and pore fluid variations, is required. Particularly because reservoir heterogeneity limits the spatial distribution of petrophysical and elastic properties used to understand the reservoir system or quantify its resource or storage potential. Traditionally, variations in reservoir rock and fluid properties are identified by interpreting depositional patterns and fluid content from core or cutting samples, or by applying statistical rock physics techniques to elastic well log data. However, collecting core data is costly, and both methods can be subjective, necessitating expert knowledge of sedimentological and rock physical principles to produce meaningful classification results. To overcome these limitations, we present a cost-effective and comparatively objective framework for identifying litho-fluid facies (LFF) using machine learning (ML) algorithms. Various statistical ML techniques were used for data-driven delineation of the LFF in the Niger Delta siliciclastic formation from a suite of commonly available geophysical well logs. The study followed a two-part process to arrive at the desired outcome. First, the target classes—clean hydrocarbon sand, shaly hydrocarbon sand, brine sand, and shale—were generated using a Dirichlet Process Gaussian Mixture Model (DPGMM) with Variational Inference (VI), an unsupervised clustering technique. These classes were subjected to probabilistic thresholding based on their log-likelihood and silhouette coefficient scores to obtain high-quality training samples. Following that, the training samples were used to construct supervised multiclass predictive models capable of generalizing the target LFF. Several classification metrics and charts were used to assess the accuracy and speed of the models to determine the model with the best predictive and computational performance. This ultimately revealed that the best-performing model was a single decision tree classifier with perfect metric scores, significantly high prediction probabilities, and minimal computational time. The random forest and gradient boosting classifiers performed similarly well on the task. Moreover, the use of analytical and statistical techniques throughout the process facilitated an objective and accurate differentiation of the rock and fluid types. The ability of the models to generalize to unseen data in a new well location with high predictive confidence makes it possible to characterize the spatially distributed facies in the study area with minimized uncertainty. Hence, we recommend the adoption of this framework for rapid and accurate LFF identification.