Insights Into Mantle Plume Geochemistry From Machine Learning

Abstract

Oceanic island basalts are the products of mantle plume melting and their chemistry provides insights into the Earth's deep interior. We report a statistical and machine learning study of 8 radiogenic isotopes and 19 incompatible trace element ratios in basalts from 27 oceanic island volcanic chains associated with mantle plumes compiled from the GEOROC and EARTHCHEM-PetDB databases. Machine-learning hierarchical analysis and agglomerative clustering results based on t-distributed stochastic neighbor embedding (t-SNE) reveal distinct clusters of isotopic compositions corresponding to canonical ones of HIMU, EM I, EM II, PREMA. and DM as well as a LOND cluster, which however do not reflect the existence of discrete components. The HIMU clan is restricted to only a couple of plumes and is characterized by low K/U, Pb/Ce, Ba/Nb and strong REE fractionation. EM I has higher K/U, Ce/Rb, Ba/Nb and lower U/Pb, Rb/Ba, Rb/Sr than EM II reflecting a difference in prevalent recycled components. Stepwise multiple regression reveals that fractionations of incompatible element ratios can be explained by variations in partial melting controlled by lithospheric thickness and plume buoyancy flux; the latter indicates that buoyancy flux primarily reflects plume temperature. ³He/⁴He also correlates with plume buoyancy flux, suggesting that the hottest plumes carry the least radiogenic He. The hottest plumes may be those rising from the core-mantle boundary. This, and the absence of evidence of a primordial mantle reservoir suggest that unradiogenic He may be derived from the core.