Applications of multivariate geochemistry to mineral deposit characterization: A case study from the Pd-mineralized Lac Des Iles Complex, western Ontario, Canada

Dimensionality reduction algorithms are increasingly being applied to high-dimensional geochemical datasets to support data interrogation and rock classification. However, their responsible implementation, particularly with closed geochemical data, necessitates careful consideration. This study applies Uniform Manifold Approximation and Projection (UMAP) to a commercially sourced whole-rock geochemical dataset from the palladium-mineralized Lac des Iles Complex (LDIC). The LDIC is an ideal case study as it comprises distinct chemical domains and numerous rock types that are overprinted by textural variation and hydrothermal alteration, making the geology difficult to systematically model and interpret. The workflow considers: (1) recalculation to anhydrous, sulfide-free compositions; (2) selection of appropriate input variables; (3) imputation of missing values; (4) transformation to mitigate closure effects; (5) tuning of algorithm hyperparameters; (6) evaluation of low-dimensional embeddings; (7) interrogation and selection of clusters. Interrogation of the UMAP embedding highlights that distinctive and often near-monomineralic lithologies ( e.g., dunite, clinopyroxenite) cluster more consistently than other rock types ( e.g., norite, gabbronorite). These distinctive lithologies are also logged more consistently than other rock types, where a comparison between derived objective clusters and lithologies logged by personnel highlights lithologies that are particularly challenging to log consistently. In cumulate mafic–ultramafic rocks, local UMAP structures reflect cumulate mineralogy, intercumulus melt proportions, and alteration, while global structures potentially relate to magmatic differentiation. As such, UMAP embeddings can help explore geochemical architecture and any lithogeochemical controls on mineralization. Objective clusters derived from UMAP embeddings can serve as categorical variables that can assist with generating bespoke discrimination diagrams, examining drill-core logs, and constructing three-dimensional geological models. Importantly, dimensionality reduction algorithms are tools to assist with data interrogation; they do not provide answers but do help the geologist to ask insightful questions. If utilized, these algorithms should be used in conjunction with petrography, lithogeochemical assessments, and geospatial analysis. • Practical applications of multivariate analysis applied to geochemical datasets • Workflow for applying the UMAP algorithm with key considerations and precautions • Using multivariate analysis to assist with unit classification and lithogeochemical interpretations