Machine learning predicts mantle endmembers from basalt trace element geochemistry

Abstract Identifying mantle endmembers is fundamental to understanding mantle evolution and heterogeneity as well as the origin of Mid-Ocean Ridge Basalts (MORBs) and Ocean Island Basalts (OIBs). The most effective tracers of mantle endmembers are isotopic ratios, which are minimally fractionated during melting and resistant to alteration; however, isotopic analyses are costly. Bivariate plots of trace elements and their ratios have also been used as mantle signatures, but are more susceptible to fractionation and alteration, limiting their diagnostic power compared to isotopes. This study aims to classify mantle components using machine learning (ML) with trace element data alone, reducing the need for costly isotopic data and overcoming the limitations of bivariate trace element plots. We compiled global geochemical data from MORBs and OIBs representing four mantle endmembers: Depleted Mantle (DM), Enriched Mantle 1 (EM1), Enriched Mantle 2 (EM2), and HIMU (high μ= 238 U/ 204 Pb). Using automated ML, we developed a supervised classification framework to assign basalts to these four endmembers based on trace element input variables alone. The AutoML best model on the full dataset achieved an accuracy of 79% on an internal test set. On 13 geographically distinct sub-regions held out from training, our AutoML pipeline achieved a pooled accuracy of 75%. These results demonstrate that trace elements contain sufficient information to partially approximate isotopic mantle endmember signatures. We provide our open-source classification tool and accompanying Python code on GitHub. To our knowledge, this is the first ML classification tool that predicts mantle endmembers using elemental data alone.