Application of deep learning to shale microstructure classification

Abstract

This study reports the application of a deep learning workflow using convolutional neural networks (CNN) to classify scanning electron microscopy (SEM) microstructural images of different shales to determine geologic formation based only on the SEM image. We used approximately 27,000 SEM images (512 × 512 pixels) from 18 different unconventional reservoir formations with a range of maturities to train a CNN through transfer learning. Our test results show a 93% accuracy in identifying the correct formation using SEM images. In addition, we also generated the probabilities of an image associating, or being similar to, different formations. These probabilities allow the user to determine what formations or zones of a formation have similar microstructures. The most important aspect of the workflow is the extremely rapid classification. After fully training the network for 120 h, we were able to predict the formation and the associated probabilities with different formations in 1.8 ms/image.