Evaluation of uniform manifold approximation and projection for dimensionality reduction of large geochemical datasets relevant to magmatic ore-fertility, petrologic classification, and provenance

Archived and curated major and trace element and isotope ratio geochemical datasets are growing at a rapid pace, commonly comprising data for many thousands of samples. The multi-dimensional nature of such vast datasets, listing information across the periodic table, is challenging to interrogate and visualize. A common pitfall is over-simplification of the analysis by focusing on ‘key’ geochemical ratios, projecting data into fields with arbitrarily set boundaries, or assigning binary classifiers. This situation invites the evaluation of available dimensionality reduction techniques. In this study, the relatively rapid non-linear dimensionality reduction technique ‘Uniform Manifold Approximation and Projection’ (UMAP) was used to simplify large tabulated (many thousands of datapoints) datasets of zircon and apatite geochemistry. The resulting two-dimensional plot reveals both local and global structure in the data. When datapoints are colour-coded for individual element concentrations, element ratios, mineral ages, or independent classification (e.g., host granite type), the geological relevance of the UMAP embedding can be evaluated. As a porphyry copper fertility prospecting tool, UMAP not only makes visual the multi-dimensional data, but the projection serves as a simplifying, pre-processing step for machine learning classification beyond strictly classifying data into ‘barren’ or ‘fertile’. Namely, the UMAP projection and point cloud structure can reveal trends with magmatic age, fertility potential vectors, and dissimilarities between regional datasets in the context of the global data structure. Beyond magmatic copper fertility, UMAP was found to group zircon according to petrologically established geochemical features, such as S-type magmatic zircon dominated by the xenotime substitution mechanism. Applied to apatite geochemistry, UMAP was able to group analyses from grains with known provenance and form data regions related to petrogenetic similarity (e.g., sedimentary and high-grade metapelite metamorphic). The overall UMAP approach of reducing dimensionality via projection into two-dimensional space and pre-processing the emerging data structure for machine learning classification can potentially be applied to a diverse range of large geochemical datasets. In doing so, care must be taken to balance local structure against global dataset coherence, whilst testing the viability of the projection and grouping through colour-coding datapoints for known geochemical parameters, age or tectonic setting. • UMAP dimensionality reduction is an effective method of visualizing large geochemical datasets, balancing local and global data structure. • Petrogenetically informed UMAP embedding highlights a zircon magmatic fertility gradient from fertile to barren. • Dataset dimensionality reduction may decrease the complexity of machine learning models needed to classify grains. • UMAP of zircon elemental compositions distinguished zircon from S-type magmas. • When applied to apatite, UMAP forms petrogenetically significant clusters.