Examining the Quantification Capability of Automated Mineralogy System: A Machine Learning Approach

Abstract Automated mineralogy (AM) systems are a novel platform for material characterization. By analyzing the X‐rays generated from sample–electron interactions, these systems can rapidly and automatically provide mineralogical and textural descriptions of a sample. However, limited mineral classification accuracy has restricted their broader application. Although various algorithms have been proposed to enhance performance, few have undergone rigorous validation. In this study, we apply a machine learning approach to systematically examine the quantification capabilities of AM systems. We first constructed a data set from three representative planetary samples—a Martian nakhlite, an asteroid angrite, and a lunar mare basalt. Then, based on the data characteristics and mission requirements, we developed a new classification algorithm. The algorithm achieved a 99.5% accuracy and good computational efficiency, confirming that AM systems can indeed provide quantitative results. The proven novelty detection capability of the algorithm helps to identify anomalous and undefined spectra, further enhancing the reliability and validity of the results. We conclude by discussing the algorithm's characteristics and potential applications, the advantages of AM systems over conventional rock characterization techniques, and the critical role of data accumulation in advancing the field.