Automated Geologic Core Description via Machine Learning

Abstract:

The task of conducting sedimentological core description is fundamental and important in building geological and simulation models. It is time consuming and subject to the geologist’s expertise. With the advancement in machine learning algorithms, it is now possible to investigate the automation of core descriptions using a convolution neural network for image recognition. In this paper, a prediction of the rock’s facies and depositional environment is tested using high-resolution images of slabbed core samples. A convolutional neural network algorithm is selected for image classification to perform a geological segmentation based on the following attributes: color, lithology, rock texture (e.g., grain size and sorting), sedimentary structure, fractures, mud content, fossils, cement and visible porosity. A training dataset composed of a total of 200 labeled core images from four different wells with full core description and interpretation for facies and depositional environment have been utilized. There are a total of 11 labels representing different depositional environments of Permo-Carboniferous Sandstones: fluvial system includes (sheetflood, paleosole, floodplain), and the Aeolian system includes dune, interdune, sandsheet, and playa facies, in addition to glacial and lacustrine depositional systems. Validation of results has been conducted through a dataset comprised of 20 core images from two additional wells. The result shows that color and sedimentary structures (e.g., laminations) are the best discriminating attributes to differentiate between the different sedimentological facies. The image quality and camera settings (light exposure and angle) are also critical to have better results.