Expert K-means reconstruction method: a novel image processing approach for mesostructure reconstruction of crystalline rocks

Abstract

Crystalline rocks exhibit pronounced heterogeneity, making the accurate reconstruction of their mesostructures a fundamental prerequisite for mesomechanical analysis. Current methods for reconstructing the mesostructures of crystalline rocks can be broadly categorized into statistical reconstruction methods and digital image processing methods. This paper systematically reviews these approaches and innovatively integrates expert systems with unsupervised machine learning, proposing the Expert K-Means Reconstruction Method (EKRM). EKRM combines the accuracy of expert systems with the objectivity of unsupervised machine learning, enabling highly precise reconstruction of rock mesostructures. Additionally, this study delves into the identification of grain boundaries in rocks, introducing a probabilistic approach to delineate mesostructural boundaries. The results demonstrate that EKRM significantly outperforms existing methods in terms of reconstruction accuracy and reusability. Furthermore, numerical simulations of the mesostructures reconstructed using EKRM were conducted and compared with laboratory experiments. The findings confirm that EKRM-reconstructed mesostructures effectively capture the influence of rock mesostructures on their mesomechanical behavior. The related code has been shared on GitHub.