Discrimination of Pb-Zn deposit types using the trace element data of galena based on deep learning.

[Display omitted] • Deep learning is used to discriminate Pb-Zn deposit types using galena trace elements. • UMAP is used to visualize deep learning internals for 'black box' insights. • SHAP is used to interpret the 'black box' nature of deep learning models. • The approach involves discrimination, visualization, evaluation, and dissection. Different types of ore deposits exhibit distinct metal sources, physicochemical conditions, and ore-forming processes. Galena, a key sulfide in Pb-Zn deposits, possesses trace elements that may be utilized for classifying Pb-Zn deposit types. Presently, research on classifying ore deposit types based on trace elements in galena is sparse, and there is a lack of robust methods for distinguishing Pb-Zn deposit types using these trace elements. In this study, we demonstrate that a deep learning algorithm, based on the trace elements in galena, can be effectively utilized for classifying Pb-Zn deposit types. The model training process utilized UMAP for visualization, and the evaluation was conducted using multiple statistical metrics, confusion matrices, and ROC curves. Finally, by dissecting the 'black box' with the UMAP algorithm, we established a comprehensive set of analysis methods for discriminating deposit types: discrimination-visualization-evaluation-dissection. A dataset comprising 828 LA-ICP-MS analyses of galena from 34 Pb-Zn deposits worldwide was curated from peer-reviewed sources. It includes data on 7 elements (Se, Ag, Cd, Sn, Sb, Tl, and Bi) across 7 deposit types: carbonate replacement deposits (CRD), epithermal, mississippi valley type (MVT), sedimentary exhalative (SEDEX), skarn, volcanogenic massive sulfide (VMS), and vein-type deposits. Initially, we selected the UMAP algorithm from three visualization methods (PCA, t-SNE and UMAP), for its ability to balance global and local structures in visualizing the internals of the deep model. Subsequently, we developed a deep learning model (1D-CNN) for deposit type classification. Various evaluation metrics and visual analyses indicate exceptional performance of our model, achieving an overall accuracy of 99.18 % on the test set. Finally, the SHAP algorithm was used to analyze the importance of trace elements in galena for different deposit types. Elements such as Sn, Tl, and Bi were found to be particularly significant in classifying deposit types, confirming the influence of multiple elements in Pb-Zn deposits within galena. Therefore, we conclude that deep learning algorithms can effectively identify Pb-Zn deposit types, offering new insights into the study of galena. [ABSTRACT FROM AUTHOR]