Traditional methods of core characterization are time-consuming and require the considerable effort of a large number of experts. At the same time, it is quite difficult to readily detect small but important elements in reservoir rocks, especially inside the core volume, that can have a strong impact on assessing reservoir potential and hydrodynamic properties. This research explores the integration of high-resolution micro-computed tomography (mu CT) imaging and transfer learning for the automated characterization of reservoir rocks and tomofacies identification. The study involved collecting and marking a dataset containing 66,560 mu CT images of 130 standard core plugs containing all major reservoir rock types. Using this dataset of 2D mu CT images of core plugs, we applied convolutional neural networks (CNNs) pre-trained on ImageNet and fine-tuned them for rock classification tasks. Applying a transfer learning approach, we trained seven ResNet-50 models, which were able to identify and classify reservoir rock features in mu CT images. These features included basic characteristics such as rock type, general texture, the presence of fractures and sulfidization that are common for all reservoir rocks. Further, depending on a specific rock type, individual features such as grain size, individual sedimentological texture and the presence of organic matter zones in the sample could be determined by the specialized model. The results show an average classification accuracy of over 94%, indicating the effectiveness of transfer learning in enhancing rock characterization. The obtained models demonstrated high validation rates and were organized into a workflow for automated rock feature characterization. A case study conducted on a carbonate core from a real oilfield provided a practical demonstration of the automated system's work, whose predictions were validated by traditional analyses such as X-ray diffraction and thin-section petrography. The produced models demonstrated the ability for rapid detection of subtle details that might otherwise have been missed during manual inspection. In addition, the concept of tomofacies resulting from the integration of mu CT images and machine learning predictions has proven useful in highlighting different reservoir zones, allowing us to understand their heterogeneity and the ability to separate rock sections despite their apparent monotony. These findings suggest that the integration of mu CT imaging and transfer learning is a promising approach for improving the automation and accuracy of reservoir rock characterization. The developed technique represents a significant advance in reservoir characterization, offering rapid and accurate identification of rock properties that can improve exploration and development strategies.
