Machine Learning and Data-Driven Classification and Prediction of Pyrite Genesis

Pyrite occurs in a wide range of ore-forming environments, and its trace element and rare earth element (REE) compositions are important geochemical indicators for determining its genetic origin, thereby providing valuable constraints for ore deposit research and mineral exploration. Machine learning, as a data-driven technique, offers new insights into the genesis types of pyrite by comprehensively analyzing data and uncovering underlying patterns. In this study, representative pyrite samples with diverse morphologies, including both sedimentary and hydrothermal types, were collected. Based on electron probe microanalysis (EPMA) and laser ablation inductively coupled plasma mass spectrometry (LA-ICP-MS), a comprehensive analysis of pyrite genesis was conducted. Furthermore, an artificial neural network (ANN) machine learning algorithm was employed to establish a classification model, and the predictive performance of this trained model was compared against that of traditional discrimination methods. The results show that when only trace elements were used as input features, the training set yielded a classification accuracy of 64.3% for sedimentary pyrite and 93.9% for hydrothermal pyrite. In the testing set, sedimentary pyrite was classified with an accuracy of 100%; however, this result may reflect the limited sample size rather than the model’s true generalization ability and therefore should be interpreted with caution. Feature importance analysis identified Cu, Zn, Te, Bi, and Pb as the key variables of this model. When both trace and rare earth elements (REEs) were used, the detection accuracy for sedimentary pyrite and hydrothermal pyrite in the training set was 87.5% and 100%. However, the combined trace + rare earth elements model exhibited a clear overfitting tendency: its overall training accuracy reached 95.2%, but its validation accuracy (88.9%) was lower than that of the trace-only model (93.8%). Feature importance analysis indicated that Pb, Ho, Zn, Ag, and Ce contribute substantially to the model. The machine learning model proposed in this study is convenient and efficient, providing a novel basis for determining the genetic types of complex pyrite.