CORE-SCALE ROCK TYPING USING CONVOLUTIONAL NEURAL NETWORKS FOR RESERVOIR CHARACTERIZATION IN THE PETROLEUM INDUSTRY

Abstract

Rock typing is an essential tool for reservoir characterization and management in the petroleum industry. It is the process of grouping portions of a rock sample based on their physical and chemical properties. This process is currently done by experts in the industry, which consumes valuable industry resources. Precise and efficient rock typing can build accurate geological models, optimize exploration and production strategies, and reduce exploration and production risks. This work proposes a deep learning method to identify and classify rocks based on their pore geometry, mineralogy, and other characteristics. The proposed technique segments a micro-CT image into different rock types using a neural network for automated rock typing. We suggest using a UNet architecture for the neural network for this task. The network has been trained in a supervised manner on expert-labelled images. The method's performance has been evaluated using qualitative and quantitative metrics. The neural network takes less than 200 milliseconds to provide the rock types, which is much faster than a human expert. We perform an explainability analysis of the neural network using class activation heatmaps approach to get insight into the learned weights. Rock typing using deep learning can improve the petroleum industry's workflow.