AI-based graptolite identification improves shale gas exploration

Abstract

Graptolites are fossils from the mid-Cambrian to lower Carboniferous periods that inform both our understanding of evolution and the exploration of shale gas [1–4]. The identification of graptolite species remains a visual task carried out by experienced taxonomists because their fine-grained morphologies and incomplete preservation in multi-fossil samples have hindered automation. Artificial intelligence (AI) holds great promise for transforming such meticulous tasks, and has already proven useful in applications ranging from animal classification to medical diagnostics [5–15]. Here we demonstrate that graptolites can be identified with taxonomist accuracy using a deep learning model [16–18]. We develop a convolutional neural network to classify macrofossils, and construct a comprehensive dataset of >34,000 images of 113 graptolite species annotated at pixel-level resolution to train the model. We validate the model’s performance by comparing its ability to identify 100 images of graptolite species that are significant for rock dating and shale gas exploration with 21 experienced taxonomists from research institutes and the shale gas industry. Our model achieves 86% and 81% accuracy when identifying the genus and species of graptolites, respectively; outperforming taxonomists in terms of accuracy, time, and generalization. By investigating the decisions made by the neural network, we further show that it can recognise fine-grained morphological details better than taxonomists. Our AI approach, providing taxonomist-level graptolite identification, can be deployed on web and mobile apps to extend graptolite identification beyond research institutes and improve the efficiency of shale gas exploration.