Automated mineralogy systems are a type of advanced material characterization platform. By scanning the surface with an electron beam and analyzing the emitted X-rays, they can automatically and efficiently extract mineralogical and morphological information from samples. However, inaccuracy in mineral classification due to instrument and sample properties limits their application. The enhancing algorithms proposed in previous studies also lack rigorous validation. In this work, we used machine learning principles and models to examine the quantification capability of the automated mineralogy instrument. We first built a well-labeled spectral dataset from a Martian meteorite, then constructed a model based on one-class support vector machine for classification. The trained model achieved a 99.4% accuracy on the test set. Additionally, it can successfully identify spectral outliers, which helps increase the classification reliability and contribute to creating more complete training sets. Thus, we confirmed that automated mineralogy systems can indeed provide quantitative results, as long as combined with robust algorithms and training data. We finally pointed out the limitations in interpretability and spatial correlations of our model, and discussed how automated mineralogy could benefit from today’s powerful artificial neural networks.