Revisiting the Geochemical Classification of Zircon Source Rocks Using a Machine Learning Approach

Abstract

Trace element fingerprints preserved in zircons offer clues to their origin and crystallization conditions. Numerous geochemical indicators have been established to evaluate the source rock characteristics from a geochemical perspective; however, multivariate trace element data have not been sufficiently investigated statistically. As substantial amounts of zircon data from a wide range of rock types have become accessible over the past few decades, it is now essential to reassess the utility of trace elements in discriminating source rock types. We employed a new zircon trace element dataset and established classification models to distinguish eight types of source rocks: igneous (acidic, intermediate, basic, kimberlite, carbonatite, and nepheline syenite), metamorphic, and hydrothermal. Whereas a conventional decision tree analysis was unable to correctly classify the new dataset, the random forest and support vector machine algorithms achieved high-precision classifications (> 80% precision, recall, and F1 score). This work confirms that trace element composition is a helpful tool for province studies and mineral exploration using detrital zircons. However, the compiled dataset with many missing values leaves room for improving the models. Trace elements, such as P and Sc, which cannot be measured by quadrupole inductively coupled plasma mass spectrometry, are vital for more accurate classification.