Lithofacies classification integrating conventional approaches and machine learning technique

Abstract

This study introduces an integrated approach for lithofacies classification utilizing core samples, wireline logs, and a machine learning technique. It is specialized for the Eagle Ford shale and the Austin Chalk where operators often face difficulties to define geological heterogeneity due to relatively consistent and low reservoir quality compared to conventional reservoirs. A set of cored slabs, thin sections, scanning electron microscope (SEM) images were utilized for lithofacies classification. Four lithofacies were defined with depositional texture, fabric, mineralogy, pore type, diagenesis, and biological features: (1) Organic-matter-rich mudstone, (2) Organic-matter-lean calcareous marl, (3) Heterogeneous argillaceous wackestone and marl, (4) Massive marly chalk. These four lithofacies are not easily classifiable through conventional wireline log cross-plotting. A convolutional neural network (CNN) model was trained to classify the lithofacies using five wireline logs (Gamma ray, bulk density, neutron porosity, deep resistivity, and compressional sonic logs) which are commonly accessible at field sites. In order to validate and test the CNN model, k-fold cross-validation and blind well test were conducted. This data-driven approach is expected to provide technical insights to operators seeking practical approaches to identifying prospective drilling locations and optimal completion strategies based on in-depth reservoir characterization.