Modeling of Cu, Pb, and Zn reservoirs in crustal igneous rocks using explainable machine learning

Abstract

Abstract Here we use explainable machine-learning (ML) analysis to determine which variables, so-called features, are most important for the prediction of trace concentrations of Cu, Pb, and Zn in a large data set for igneous rocks. After data filtering and preparation, a gradient-boosted decision tree regression model is trained using eXtreme Gradient Boosting (XGBoost). Feature importance was evaluated quantitatively for each of the trace elements. In addition, we performed SHapley Additive exPlanations (SHAP) analysis to test the robustness of the feature importance results. Simple calculation of historical and known reserves shows that < 0.1% of the weight fractions of Cu, Pb, and Zn are contained in the ore deposits. The rest must be dispersed in rocks. For Pb, the most important features, determined by ML, are K 2 O, Rb, U, and Th. They are interpreted in the sense that Pb is accumulated in the rest magma during fractionation and, after solidification, stored mostly in K-feldspars and in minerals with primary elevated content of U or Th, such as allanite or zircon. For Zn, the most important features are TiO 2 , MnO, Ga, and Sc. They are assigned to ilmenite, a lesser extent magnetite, and amphiboles/pyroxenes, as the Zn reservoirs. These interpretations are supported by a compilation of known distribution coefficients. For Zn, one of the important features is also Pb, and vice versa. This relationship can be due to post-magmatic, low-temperature alteration and weak disseminated mineralization with Pb and Zn sulfides. Other important features for Zn are SiO 2 and Zr, indicating a relationship between the degree of igneous evolution and Zn content. For Cu, the only important feature by far is V. Such relationship could be explained by the control of Cu concentration through the oxidation state of magmas that also influences the redox behavior and partitioning of V. The identification of the reservoirs of As, Sb, and Bi failed; perhaps these elements are stored in sulfides. In the present work, the focus was not on predictive accuracy, but rather on data driven feature importance analysis that revealed systematic trends in the dominant geochemical controls for the three elements targeted. These findings can be extended for further data-driven assessment of trace-element in igneous and other geological systems.