Abstract
Zircon is a common, highly refractory accessory mineral found in sedimentary, igneous and metamorphic rocks that forms through various processes (e.g., magmatic, hydrothermal, metamorphic growth and recrystallization). Trace element compositions of detrital zircon have been widely applied to understanding tectonic processes and continental reconstruction. Previous studies have shown that no single compositional parameter (i.e., rare earth elements [REE] pattern and Th/U) can distinguish zircon petrogenesis. Here, we employ novel machine learning methods to differentiate between 'metamorphic-like' and 'igneous-like' zircon, independent of researcher-labelled lithologies and categorized by their dominant REE signatures. With advancements and the adoption of machine learning algorithms by the geochemical community, we holistically analyse n-dimensional geochemical data (14 dimensions for the full REE suite) to ascertain or confirm complex geological processes. We utilize a dataset of previously identified zircon with varying proposed origins (i.e., metamorphic, igneous, S-type granites and early Earth detrital) and unsupervised machine learning algorithms (i.e., UMAP and Mclust), which reveal two distinct clusters that we interpret as the dominant magmatic-like and metamorphic-like REE signatures. The two REE signature clusters are used to define a comprehensive REE equation that matches an unknown zircon into the most similar interpreted cluster (magmatic-like or metamorphic-like) through probabilistic bimodal logistic regression. The logistic regression equation was derived exclusively from REE signatures independent of previous classification. Model validation was achieved using confusion matrices and yielded a mean accuracy of 0.91 (balanced accuracy = 0.88), with average precision = 0.90, recall = 0.80, F1 = 0.85 and AUC = 0.96, indicating robust discriminative performance. Unsupervised machine learning demonstrates that two distinct REE signatures emerge, which vary in their HREE enrichment and LREE depletion. We demonstrate that machine learning can decipher complex multi-elemental relationships in zircon REE data and can be used to classify zircon.