Machine learning based microstructure classification using SEM images and morphological features

Abstract This study implements an approach based on segmentation-based morphological features derived from scanning electron microscopy images. Microstructures were classified as heterogeneous, homogeneous, and partially heterogeneous using physical features extracted through Otsu-based segmentation and morphological operations. The dataset was expanded to 2000 synthetic images using Poisson-based patch synthesis. The performance of Random Forest, K-nearest neighbors, Support Vector Machine, and Extreme Gradient Boosting algorithms was evaluated comparatively. Results demonstrate that the Random Forest model achieved the highest performance with an accuracy of 87.65 % and a Macro-F1 score of 87.83 %. Confusion matrices reveal that the primary source of error across all models is the transition structures. The findings indicate that features provide a robust and interpretable representation for microstructure characterization. This study shows that machine learning approaches based on physical features can produce reliable and generalizable results in microstructure classification problems, proving effective for explainable artificial intelligence applications in the field of materials science.