Machine learning-driven geochemical fingerprinting and risk characterization of mineral dust across different...

Investigating mineral dust emitted from mining activities enables the assessment of environmental risks posed by potentially toxic elements (PTEs) and the discrimination of geochemical fingerprints characteristic of distinct operational settings. Accordingly, this study employed site-specific dust sampling, geochemical analysis of PTEs using ICP-AES, supervised machine learning (e.g., Support Vector Machine and Multinomial Logistic Regression), multivariate statistics (e.g., Principal Component Analysis), pollution and ecological indices (e.g., Pollution Load Index), and health risk modeling to delineate PTE contamination patterns, determine high-risk microenvironments, and identify geochemical fingerprints (e.g., ore-handling zones vs. confined cabins) within El-Gedida Iron Mine (Western Desert, Egypt), thereby establishing dust-borne elemental profiles as tracers for evidence-based environmental intervention. Mean PTE concentrations decreased in the order of Fe > Mn > Zn > Cr > Pb > Cu > Ni, with Cu showing extreme variability (CV = 142.6%) and a 40-fold range, linked to a localized enrichment. Composite indices exhibited substantial contamination across all samples, with a mean PLI of 2.21. Cr and Ni posed unacceptable lifetime cancer risks in children (TCR = 6.87E-04 and 2.28E-04, respectively), while Cr exhibited the highest non-carcinogenic risk (HI = 0.522), though below the critical threshold (HI < 1). Supervised machine learning models demonstrated reliable group separability and probabilistic discrimination driven by key elemental predictors (e.g., Cu), effectively extracting latent geochemical signatures, with prominent examples including the Cu-Pb-enriched fingerprint indicative of confined drilling cabins, reflecting localized accumulation from internal vehicular emissions, and the Fe-Mn lithogenic-derived signature characteristic of ore-handling zones. The Multinomial Logistic Regression (MLR) model achieved a predictive accuracy of 95.8%, highlighting the framework's strong practical applicability.