Machine learning models to predict rare earth elements distribution in Tethyan phosphate ore deposits: Geochemical and depositional environment implications

Abstract

The global market for rare earth elements (REE) is growing rapidly, driven by rising demand and limited production sources, prompting interest in recovering REE from secondary sources such as phosphate deposits. The Tethyan belt, extending across North Africa and the Middle East contains substantial Upper Cretaceous to Eocene REE-rich phosphorite deposits but with limited geochemical data available. This study provides a novel machine-learning (ML) method to predict REE contents in these deposits and verify a useful geochemical classification based on the concentrations of nine major element oxides. Four ML models are developed to achieve this: eXtreme Gradient Boosting (XGBoost), Random Forest (RF), Support Vector Regression (SVR), and Decision Tree (DT). The datasets are divided geochemically into oxic and sub-oxic patterns and these are evaluated with the ML models separately and in combination to accurately predict light REE (LREE), heavy REE (HREE), and total REE contents (∑REE). For the oxic pattern dataset, Fe₂O₃ and K₂O exhibit the highest feature importance consistent with a glauconite influence. For the sub-oxic pattern dataset, MnO and SiO₂ exhibit the highest feature importance consistent with high terrigenous inputs (MnO), and silicification. The ML results support the importance of the local deposition environment in determining REE distributions in these deposits. Paleogeography, ocean-margin tectonics, sea-level oscillations, and marine currents exert influence on the local depositional environments. The eXtreme Gradient Boosting model generates the lowest REE prediction errors for all the datasets evaluated.