A machine learning method for distinguishing detrital zircon provenance

Abstract

Zircon geochemistry provides a sensitive monitor of its parental magma composition. However, due to the complexity of the uptake of trace elements during zircon growth, identifying source magmas remains challenging, particularly for detrital grains whose petrological context is lost. We use a machine learning-based approach to explore the classifiers for zircon provenance, based on 3794 published, high-quality zircon trace element analyses compiled from I-, S-, and A-type granites. Three supervised machine learning algorithms, namely, Support Vector Machine (SVM), Random Forest (RF), and Multilayer Perceptron (MLP) were used and trained with 11 features, including 7 trace elements (Ce, Eu, Ho, Nb, Ta, Th, and U) and 4 derived trace element ratios (Th/U, U/Yb, Ce/Ce*, and Eu/Eu*). Our results show that all three trained machine learning methods perform very well with accuracy varying from 0.86 to 0.89, and that input–output relationships captured by different ML methods are nearly consistent and can be explained by the known petrological processes. The application of our trained machine learning classifiers to detrital zircon studies will enhance the interpretability of zircon assemblages of different origins. It also helps develop interpretations, approaches, and tools that will benefit, for example, the study of continental crust evolution and mineral exploration.