172-1 Discriminating mine tailings from different ore deposits using supervised machine learning

Session: Environmental Geochemistry and Health (Part II)

Presenting Author:

Mouataz Mostafa

Authors:

Mostafa, Mouataz T. ¹, Abd El-Bakey, Sahar M. ², Elwam, Aya M. ³, Zakaria, Mariam Y. ⁴, Abu Elwafa, Reham Y.⁵, Farhat, Hassan I. ⁶

(1) Geology Department, Faculty of Science, Arish University, Arish, North Sinai, Egypt, Arab Rep., (2) Department of Geological and Biological Sciences, Faculty of Education, Ain Shams University, Ain Shams, Cairo, Egypt, Arab Rep., (3) Geology Department, Faculty of Science, Suez Canal University, Ismailia, Ismailia, Egypt, Arab Rep., (4) Geological and Biological Department, Faculty of Education, Ain Shams University, Ain Shams, Cairo, Egypt, Arab Rep., (5) Geology Department, Faculty of Science, Sohag University, Sohag, Sohag, Egypt, Arab Rep., (6) Geology Department,, Faculty of Science, Suez University, Suez, Suez, Egypt, Arab Rep.,

Abstract:

This study evaluated the capability of supervised machine learning (ML) models to discriminate the geochemical signatures of mine tailings derived from different ore deposits, providing insight into the key discriminant potentially toxic elements (PTEs) that characterize each tailings type, which guides targeted environmental monitoring and mitigation strategies, particularly within multi-mine districts. To this end, 126 tailings samples representing six ore types (phosphate, manganese, iron, fluorite, coal, and zinc) were collected from Egyptian mining districts and subjected to geochemical analysis before calculating pollution indices and applying ML-based classification. Mn tailings exhibited the highest As concentrations (mean = 16.85 mg/kg), whereas Cd was mainly associated with Zn and phosphate tailings, averaging 42.57 and 8.99 mg/kg, respectively. Pb in Zn tailings exhibited extremely severe contamination (CF_mean = 107.3) with a corresponding very high ecological risk (Er_mean = 536.5). Considering ML models, Partial Least Squares Discriminant Analysis (PLS-DA) identified Ni (VIP = 1.78), Fe (VIP = 1.24), and Cr (VIP = 1.13) as the principal discriminant variables (VIP > 1), highlighting their highest contribution in distinguishing tailing types. Notably, the Decision Tree model misclassified only three Mn-derived tailings samples as Fe-related material, achieving an overall test accuracy of 97.62% (misclassification rate of 2.38%), with a relative AUC values of 1.0 across all classes, whereas the multinomial logistic regression (MLR) achieved 94% classification accuracy (likelihood ratio test: χ² = 150.51, df = 45, p = 2.69 × 10⁻¹³).  Furthermore, the RBF-kernel Support Vector Machine (SVM) (C = 1, γ = scale) effectively captured nonlinear relationships among nine PTE predictors, with stable accuracy across 5-fold cross-validation. Ultimately, this high classification capability qualifies supervised ML models for reliable source attribution, supporting more targeted environmental monitoring in complex mining districts.

© Copyright 2026 The Geological Society of America (GSA), all rights reserved.