부산대학교 도서관

Petrographic observations are vital for carbonate pore‐typing, linking geological frameworks to petrophysical behavior. However, current petrographic pore typing is manual, with the qualitative to semi‐quantitative results not easily fitted into quantitative subsurface characterization. Some recent studies have automated this process using supervised machine learning (ML) and deep learning (DL), focusing on simple pore morphological features, and have reported high classification accuracies for several complex pore types. However, there are concerns about the validity of these studies due to conceptual and technical flaws in their collective approach. This study was aimed at a more fundamental problem, classifying between open microfractures and open pores in petrographic thin sections using an object‐based approach and explainable supervised ML. We analyzed 18 carbonate thin sections from the USA, numerically representing them using five shape features: compactness, aspect ratio, extent, solidity, and formfactor. Using a labeled data set of 400 microfractures and 400 pores, we evaluated nine of the most widely used supervised models. All models showed high testing accuracies (89.58%–90.42%). Interestingly, complex non‐linear models did not significantly outperform simpler linear ones. Compactness and aspect ratio were the most informative features. However, the labeled data sets did not reflect the overall data set's complexity, which suggested that high accuracies in similar studies might be due to curated data sets rather than accounting for the true complexity of carbonate pore systems. The study concludes that simple shape features are ineffective for classifying carbonate pore types. It is hoped that this study will provide a foundation for more robust artificial intelligence‐assisted pore typing. Plain Language Summary: Carbonate pore‐typing is a critical task for determining rock types. Petrographic pore‐typing from thin sections is the most mature form of carbonate pore‐typing and is vital in relating the geology of the studied formations to its petrophysical properties. To date, this process has remained manual, bound by human limitations, and difficult to link to quantitative digital reservoir models. Recent research has tried to automate petrographic pore‐typing using machine learning (ML) and deep learning (DL), claiming very high accuracies. However, there are concerns about these claims due to potential flaws in the methods used. There is potential in using ML for binary classification, especially when distinguishing between microfractures and pores, as they are quite distinct in shape. In this study we used an object‐based, supervised ML approach to differentiate these two classes, using data from 18 carbonate thin sections sourced from the USA. The data was represented using five popular shape features: namely, compactness, aspect ratio, extent, solidity, and formfactor. We used nine popular linear and non‐linear supervised ML models. The ML models tested had an accuracy of around 90 percent. Interestingly, the more complex non‐linear models did not perform much better than simpler, linear models, suggesting that distinguishing between microfractures and pores might be a straightforward problem. Among the shape features, compactness and aspect ratio proved the most useful in separating the two classes. However, we also report that the labeled data set used for training the models did not represent the full data set well, thus indicating that simple shape features cannot accurately capture the complexity of carbonate pore types even at the base binary level. The study concludes that while ML is promising for simplistic data sets, we must consider more complex shape features and build larger data sets to develop DL models. The hope is that this research will guide future efforts in ML and DL approaches to carbonate pore‐type classification. Key Points: The first study to propose a binary framing for machine learning driven petrographic pore typingLinear and non‐linear models perform equally well for idealized microfractures and poresWe highlight the need for greater scrutiny in artificial intelligence models for petrographic pore typing [ABSTRACT FROM AUTHOR]