Abstract Efficient processing of complex ores is often hampered by an incomplete understanding of the spatial distribution of critical mineralogical properties. Here, we introduce a machine learning-based framework that addresses this gap. Using the multivariate adaptive regression splines (MARSpline) method, we develop predictive models for converting standard geochemical assays into mineralogical properties – a process we term Element-to-Mineral-Properties Conversion (EMPC). Applied to the Koashva apatite-nepheline deposit (Kola Peninsula, Russia), our approach yields reliable predictions (r typically > 0.85 for major minerals and > 0.6 for key trace elements in apatite) for mineral composition and ore mineral chemistry using only bulk-rock P 2 O 5 , Al 2 O 3 , and TiO 2 . These predictions enable the construction of the first 3D block models of mineral composition and mineral chemistry for this deposit. By integrating these models with conditional geometallurgical criteria, we identify and spatially delineate zones of refractory ore (containing > 6 vol% pyroxenes and > 1 vol% “liebenerite”) and areas yielding low-grade apatite concentrate (where apatite contains > 6.5 wt% SrO + REE 2 O 3 ). This work establishes a generalizable, data-driven pathway for geometallurgical modeling, transforming historical exploration data into a critical decision-making tool for optimizing mine planning and mitigating processing challenges.
Kalashnikov et al. (2026) studied this question.