Application of Machine Learning to Characterizing Magma Fertility in Porphyry Cu Deposits

Large and easily accessible porphyry Cu deposits have already been identified, exploited, and gradually exhausted. Novel strategies, therefore, are required to identify new, deeply buried deposits. Previous studies have proposed several lithogeochemical and mineralogical approaches for identifying porphyry Cu systems. Most of these methods, however, require significant a priori knowledge of the exploration region and are, generally, of low effectiveness. In this study, machine learning models using Random Forest and Deep Neural Network algorithms are utilized to characterize magma fertility. The two models have first been trained using a large trace‐element data set of magmatic zircon and then validated on unseen data set during the training process. The performance of both models was evaluated using a fivefold cross‐validation technique, which demonstrates that the models provide consistent results and yield good classification accuracy (∼90% on average) with low false positive rates. Feature importance analysis of the models suggests that Eu/Eu*, Eu/Eu*/Y, Ce/Nd, Ce/Ce*, Dy, Hf, and Ti are the important parameters that distinguish fertile and barren zircons. The real‐world applicability of the validated models was evaluated using two well‐characterized porphyry Cu deposits in subduction and postcollisional settings—the Highland Valley porphyry Cu district (south‐central British Columbia, Canada) and the southern Gangdese belt (Tibet, China), respectively. The results demonstrate that our generalized models can discriminate zircon from igneous rocks associated with porphyry Cu deposits from those in nonmineralized systems with high accuracy and independent of geological setting, suggesting that this new approach can be used effectively in greenfield and brownfield exploration. Plain Language Summary Mineral resources, which are important to the development of human society, are increasingly being consumed as our need for technological advancement increases. With this consumption comes the need to identify additional resources, many of which may be invisible at the Earth’s surface. This challenges current exploration methods, requiring novel techniques to be developed to identify variably mineralized rock below Earth’s surface. In this study, machine learning methods are trained using a global data set of zircon chemistry to evaluate the prospectivity of porphyry Cu mineralization in magmatic districts. This work demonstrates that machine learning methods