Machine Learning for Identification of Primary Water Concentrations in Mantle Pyroxene

• Published in: Geophysical research letters 2021-09, Vol.48 (18), p.n/a
• Main Authors: Chen, Huan, Su, Cheng, Tang, Yong‐Quan, Li, An‐Zhou, Wu, Sen‐Sen, Xia, Qun‐Ke, ZhangZhou, J.
• Format: Article
• Language: English
• Subjects: clinopyroxene phenocryst; Composition; Diffusion; Dimensional analysis; Earth mantle; Estimates; Evolution; Geochemistry; Hydrogen; hydrogen diffusion; Hydrologic cycle; Hydrological cycle; Lava; Learning algorithms; Machine learning; Magma; mantle water content; Methods; Modelling; Moisture content; support vector machine; Support vector machines; Water content
• ISSN: 0094-8276; 1944-8007
• DOI: 10.1029/2021GL095191

The approach of estimating the H2O content of basaltic magmas via clinopyroxene (cpx) phenocrysts is a potentially effective way to glimpse the deep Earth water cycle. However, it is difficult to ascertain using traditional geochemical methods whether hydrogen (H) measured in cpx phenocrysts represents a primary signature that can ultimately inform estimates of the mantle water content. In this study, we conducted machine learning on the major element compositions and H2O content of cpx phenocrysts (1904 samples in total). Using the support vector machine (SVM), we defined a classifier (overall accuracy >92%) that can separate cpx that have undergone H diffusion, and thus modification of their original water content, from those that have not experienced H diffusion. Our trained SVM model has broad implications for understanding the primary water content of magma, the variations in water content during magma evolution, and the water cycle in the deep Earth. Plain Language Summary The H2O content of clinopyroxene (cpx) phenocrysts in basaltic magma has become an important way to evaluate the H2O content of Earth's mantle. However, it remains controversial whether samples collected in the field can truly reflect mantle water contents, particularly given the propensity for hydrogen to diffuse out of the cpx lattice during magma evolution. The support vector machine (SVM) is a powerful and mature machine learning method that can classify samples into one of two groups based on the analysis of high‐dimensional datasets. We have applied SVM to separate cpx into examples that either have, or have not, undergone H‐diffusion, using a database of 1904 known samples. Each sample represents an in‐situ analysis of H2O content and major element composition on a cpx grain. The trained SVM model is effective in distinguishing whether individual cpx phencrysts have preserved their initial H2O content. Moreover, when applied to a suite of well characterized samples, the model predictions are consistent with independent assessments of H‐diffusion. As a complement to traditional geochemical methods, our SVM model is a novel approach with broad implications for understanding variations of water content in magmatic systems and refining estimates of the water content of Earth's mantle. Key Points A machine learning model for judging H diffusion was built by learning of the data on the H2O content of cpx phenocrysts in mafic rocks The trained model can effectively distinguish whet