Classification of phases in alkali-activated glass powder using machine learning with Gaussian mixture model

Traditional methods for classifying micromechanical properties in alkali-activated materials depend on manual correlation of nanoindentation data, which is both time-consuming and subjective. This study examines the application of unsupervised machine learning to automate phase identification in alkali-activated glass powder and blast furnace slag. Grid nanoindentation was combined with scanning electron microscopy and energy dispersive X-ray spectroscopy to uncover heterogeneous phase assemblages. A Gaussian mixture model (GMM) was utilized to distinguish among the outer matrices, particles, rims, and their respective proportions. The GMM-based results were compared with those obtained through manual classification. The optimal number of clusters was determined using the Bayesian information criterion. Accuracy was assessed based on phase prediction error and normalized center prediction error. The tied covariance model with eight clusters showed the highest agreement with manually classified phases, which minimizes centroid and phase fraction errors. This approach enables robust, quantitative evaluation of micromechanical properties in glass-based phases, significantly reducing the need for manual classification.