Unsupervised geochemical characterisation of deeply weathered terrains and regolith-hosted REE deposits: Rationale and benefits for exploration

Abstract

The accurate characterisation of regolith materials is crucial for mineral exploration, yet distinguishing visually indistinct clay-rich samples can be challenging and labour-intensive. This study conducts unsupervised k-means clustering and principal component analysis (PCA) on a geochemical dataset of over 3000 regolith samples from the Splinter Rock rare earth element (REE) prospect, Western Australia, to determine how unsupervised statistical methods may expedite the characterisation of regolith samples across large, buried and/or regolith-hosted ore deposits. K-means clustering identified five laterally consistent regolith horizons at Splinter Rock, which were manually interpreted into three REE-barren transported horizons and two mineralized saprolite-saprock horizons. The mineralogical and metallurgical features of all 3000 samples were then extrapolated from hyperspectral and metallurgical data of a select few reference samples within their clusters, to provide a preliminary understanding of the deposit’s overall structure and properties. Despite being a first-order approach, this method highlighted several consistent, statistically robust and previously unidentified patterns across the entire prospect: 1) the highest REE grades exist predominantly in the granitic saprolite and saprock; 2) relative to the light REEs (La–Sm), the heavy REEs (Eu–Lu) experience enrichment at the saprolite-saprock boundary and depletion with increasing depth in the saprock; 3) optimal metallurgical conditions occur near this saprolite-saprock interface; 4) relative accumulation of the economically- and environmentally-important ‘magnet’ REEs (MagREE, Pr, Nd, Tb, Dy) occurs mostly in the saprock; and 5) relative MagREE enrichment can be linked to the formation of negative Ce anomalies at lower stratigraphic positions. Lastly, PCA facilitated the development of tailored geochemical ratios to classify future samples into their appropriate horizons. This study highlights unsupervised statistical analysis of existing geochemical data as a robust, rapid and effective first-pass method for classifying and characterising extensive sets of regolith samples, as well as an efficient method of outlining deposit-scale trends and zones of consistent economic REE enrichment in large regolith-hosted deposits/prospects.