Abstract: Recently, it has become an important part of petrological and geochemical research including the field of geodynamics to identify tectonic discrimination in magmatic rocks formation. Traditionally, basalt is considered the best member of rocks to identify tectonic settings. However, the use of clinopyxene has not been effective in this aspect due to limitations in data usage; and in theory gabbro is not consistent in its characterization under different tectonic environments. For this reason, it is imperative to explore the use of machine learning algorithms in studying global clinopyroxene gabbro. Here, we mainly focus on feature screening and data classification of clinopyxene gabbros from three structural backgrounds: island arc (IAB), oceanic island (OIB) and mid-ocean ridge (MORB). From the GEOROC database, after data collection and processing, we identified 385 entries of island arc, 756 ocean island and 5500 mid-ocean clinopyxene gabbros. Most data were of main elements and the rest trace elements. During feature extraction, we used chi-square test to judge feature independence, F-test to estimate the linear dependence between two random variables, and mutual information method to capture other kinds of statistical correlations. Statistically reliable data features were obtained using these three methods. During data classification, we compared the performances of three mainstream machine learning classification algorithms, namely K-Nearest Neighbor, Decision Tree and Support Vector Machine on gabbro data. The results show that for clinopyxene gabbros in above three tectonic settings, Al2O3 and TiO2 were the most distinctive main elemental compositions in clinopyroxene gabbro for differentiation, while Sr was the most discriminating trace element. The backgrounds classification accuracy of the machine learning model on data of main and trace element both reached 94%.

Key words: clinopyroxene, gabbro, island arc, oceanic island, mid-ocean ridge, machine learning