Abstract Objectives/Scope This study aims to develop and validate a fully automated AI-based thin section interpretation model to support reservoir characterization in the SE Abu Dhabi reservoirs. By targeting key carbonate formations - Shuaiba, Kharaib, Lekhwair and Habshan - the objective is to reduce manual interpretation time while enhancing the spatial extrapolation of rock properties. These AI models are designed to assist in field development workflows by providing consistent, scalable insights from petrographic data, ultimately improving the resolution and efficiency of reservoir modeling. Method A method for AI-assisted thin section interpretation was developed using advanced deep learning segmentation models and object detection algorithms, all reinforced with geological rules. This solution combines classical regression models and deep learning techniques to provide geologists with a detailed set of reservoir properties derived from thin-section images. The approach includes training supervised convolutional neural networks, establishing consistent labeling procedures, and removing image artifacts from both input and output. Additionally, continuous communication with subject matter experts ensures that predictions remain geologically realistic and help supplement limited training data. In this study, selected thin sections collected from the Shuaiba, Kharaib, Lekhwair and Habshan Formations in an oil field in the SE of Abu Dhabi were analyzed through both standard petrographic observations and the proposed AI-assisted workflow to be able to compare and contrast the obtained results. Result The used tool demonstrated an overall strong degree of accuracy when detecting and identifying grains, pore types and abundance, cement features, MICP estimations, depositional environments and textures. Its main limitations appeared to be related to instances of high facies complexity and heterogeneity, or to limited training datasets, which resulted in lower accuracies. However, the tool also provided relatively accurate diagenesis and permeability predictions, which further assisted in achieving a reliable categorization of the analyzed samples into the existing rock typing schemes available for the studied formations. The main outcome of this integration of domain knowledge and data science was an improvement in key evaluation metrics, such as intersection-over-union (IoU), precision, recall, and R2. In addition to these performance gains, the approach ensured that fundamental geological principles were maintained. The algorithm constrained petrographic object detection using biostatistical population criteria, preventing non-natural combinations of nested framework grains and ensuring geologically realistic predictions. Novel/Additive Information The application of the proposed AI-assisted petrographic interpretation workflow demonstrated a high potential for streamlining and enhancing the routine analysis of large thin section databases, which often lie underutilized in storage. These enhancements can also be subsequently implemented and deployed at a wider company and even inter-company scale, allowing specialists to conduct and coordinate their geological work through conventional web-browser applications.
Pal et al. (Mon,) studied this question.