Identification of Lithology from Well Log Data Using Machine Learning

Abstract

INTRODUCTION: Reservoir characterisation and geomechanical modelling benefit significantly from diverse machine learning techniques, addressing complexities inherent in subsurface information. Accurate lithology identification is pivotal, furnishing crucial insights into subsurface geological formations. Lithology is pivotal in appraising hydrocarbon accumulation potential and optimising drilling strategies.
 OBJECTIVES: This study employs multiple machine learning models to discern lithology from the well-log data of the Volve Field.
 METHODS: The well log data of the Volve field comprises of 10,220 data points with diverse features influencing the target variable, lithology. The dataset encompasses four primary lithologies—sandstone, limestone, marl, and claystone—constituting a complex subsurface stratum. Lithology identification is framed as a classification problem, and four distinct ML algorithms are deployed to train and assess the models, partitioning the dataset into a 7:3 ratio for training and testing, respectively.
 RESULTS: The resulting confusion matrix indicates a close alignment between predicted and true labels. While all algorithms exhibit favourable performance, the decision tree algorithm demonstrates the highest efficacy, yielding an exceptional overall accuracy of 0.98.
 CONCLUSION: Notably, this model's training spans diverse wells within the same basin, showcasing its capability to predict lithology within intricate strata. Additionally, its robustness positions it as a potential tool for identifying other properties of rock formations.