Vision Transformers based Pore Type Classification for Carbonate Reservoir Characterization

Summary Classifying pores into their intrinsic depo-diagenetic or petrophysical morphotypes is a fundamental practice within carbonate petrography, typically performed qualitatively/semi-quantitatively. In this study, we present a framework that shifts the scope of pore typing towards quantitative and reproducible data analytics by developing a Vision Transformers (ViT) based model to classify macropores in thin sections into their respective Lucia classes. The ViT architecture is augmented by the addition of a Feature Fusion block, which integrates spatial features from a Convolutional Neural Network (CNN) with ViT features, enhanced by a Global Token Addition (GTA). Critically, our classifier leverages neighbourhood information based upon the recognition that pore types cannot be solely classified based upon shape, but also requires the consideration of the pore system’s local spatial context. Trained and tested using 4,115 labelled pores obtained from 25 high-resolution scans, our model provides an accurate, automated classification of carbonate macropores, demonstrating higher precision and recall rates than benchmark CNN models. The high throughput pore-textural classification capabilities demonstrated herein offer unprecedented opportunities in the quantitative characterization of carbonates, providing improved scalability of petrographic datasets utilized for rock typing, or for studies aiming to elicit broad spatiotemporal trends in carbonate petrophysics and diagenesis.