Machine Learning for Characterizing Magma Fertility in Porphyry Copper Deposits: A Case Study of Southeastern Tibet

Abstract

Numerous intermediate to felsic igneous rocks are present in both subduction and collisional orogens. However, porphyry copper deposits (PCDs) are comparatively rare. The underlying factors that differentiate fertile magmas, which give rise to PCDs, from barren magmas in a specific geological setting are not well understood. In this study, three supervised machine learning algorithms: random forest (RF), logistic regression (LR) and support vector machine (SVM) were employed to classify metallogenic fertility in southeastern Tibet, Sanjiang orogenic belt, based on whole-rock trace element and Sr-Nd isotopic ratios. The performance of the RF model is better than LR and SVM models. Feature importance analysis of the models reveals that the concentration of Y, Eu, and Th, along with Sr-Nd isotope compositions are crucial variables in distinguishing fertile and barren samples. However, when the optimized models were applied to predict the datasets of Miocene Gangdese porphyry copper belt and Jurassic Gangdese arc representing collision and subduction settings respectively, a marked decline in metrics occurred in all three models, particularly on the subduction dataset. This substantial decrease indicates the compositional characteristics of intrusions across different tectonic settings could be diverse in a multidimensional space, highlighting the complex interplay of geological factors influencing PCD's formation.