Enhanced classification of pyrite generations based on mineral chemistry using uniform manifold approximation and projection (UMAP)

Trace element signatures are tracked to unravel the genetic history of ore deposits in several mineral systems. This is particularly true for pyrite, a ubiquitous component of many ore-forming systems, including orogenic gold deposits. Here, we critically compare the efficacy of utilizing Uniform Manifold Approximation and Projection (UMAP) versus Principal Component Analysis (PCA) as dimensionality reduction tools applied to a 31 element dataset collected using Laser Ablation Inductively Coupled Mass Spectrometry of pyrite grains from the Kibali gold district (Democratic Republic of Congo). Because of the non-linearity inherent to mineral chemistry and because of its superior preservation of local (distances within clusters) and global (separation between clusters) data relationships, the UMAP approach outperforms dimensionality reduction by PCA. We further present a workflow in which UMAP dimensionality reduction is followed by k-means clustering to guide the classification of pyrite generations in the Kibali case study. Validating this approach with trace elements and UMAP + k-means on a seed-by-seed petrography basis shows that the workflow significantly enhances the original pyrite classification, previously based on texture. This study thus emphasizes the utility of employing advanced statistical analysis methods to capture the intricate nature of pyrite formation. These findings will shape best practices for handling large multi-element datasets in pyrite mineral chemistry studies and are extrapolatable to other mineral systems in which trace element signatures are used to infer the conditions of ore deposit genesis.