Self-supervised vision transformers for classification in unconventional shales

Abstract Deep learning has significantly advanced image segmentation and classification, with convolutional neural networks (CNNs) traditionally dominating due to their translation and rotation invariance. However, vision transformers (ViTs) are emerging as powerful alternatives, particularly for capturing long- and short-range dependencies without requiring large, labeled data sets. These labeled data sets are time- and resource-intensive to acquire and are the major drawback of CNN-based approaches. In this study, a ViT-based architecture was applied to classify scanning electron microscopy (SEM) images from 22 unconventional shale plays across North and South America and Europe. Our data set comprised 22,440 grayscale backscattered electron SEM images. Our goal was to identify the source of the image from among 22 different unconventional plays. A reasonable classifier should possess a low rate of misclassification; however, persistent misclassification of an image from one specific play for another indicates microstructural similarities that can be exploited for operational purposes. A classifier such as the one developed in this paper can be used on any grayscale SEM image to identify a digital analog. Our approach was fully self-supervised. Using a teacher–student ViT framework, the model learned progressively without human-annotated labels. Over time, the networks distilled patterns from the data, learning distinguishing features between classes. This was then exploited by pairing the ViT-based algorithm with a k-nearest neighbors (k-NN) classifier that groups the images. These groups, remarkably, corresponded to the source of the images, and an accuracy exceeding 95% was consistently achieved. Remarkably, without prior training on labeled data, the model successfully identified key image features that inform its classification through attention maps. Our study underscores the remarkable ability of ViTs embedded in a student–teacher framework to learn underlying features and enable rapid, self-supervised classification.