NeutroChem Fusion: an uncertainty-aware source diagnosis framework for mining impacted waters in West Africa

Abstract

Mining impacted waters in West Africa are mixed chemical systems in which mining-derived Hg-Pb-Cd signals, geogenic As-Fe-Mn mobilisation, acid mine drainage, agricultural runoff, domestic inputs and sediment remobilisation may coexist. Existing water-quality indices, receptor models, geochemical models and machine-learning classifiers are useful for screening and apportionment, but they may compress overlapping mechanisms into dominant factors or exclusive labels. This study develops NeutroChem Fusion as a neutrosophic-Bayesian constrained DSmT framework for uncertainty-aware source-mechanism diagnosis and treatment-pathway decision support. The framework was internally evaluated on 720 field-water-system records from Ghana, Burkina Faso, Mali, Guinea, Côte d’Ivoire, Senegal and Niger sampled from January to May 2026. The full model showed source-label concordance of 0.9023 in the internal test set and 0.8730 in the workbook-held-out country-transfer subset. Pooled fixed-validation concordance was 0.8880 (95% bootstrap interval 0.8494-0.9228) and pooled macro-F1 was 0.8786. On the same fixed split and field/reference labels, multinomial logistic regression, PCA-logistic regression and random forest achieved concordances of 0.7568, 0.7606 and 0.8147, respectively; these comparisons are internal supervised benchmarks rather than proof of external superiority. The engineering chemistry layer was experimentally calibrated using standardised bench-scale adsorption, precipitation, electrocoagulation and membrane tests. The variables biochar_pct, iron_oxide_adsorption_pct, lime_precipitation_pct, electrocoagulation_pct and membrane_rejection_pct are measured removal or rejection percentages, while the associated 0-1 process scores are normalised indices calibrated against those experiments. Experimental treatment outcomes for internal test and workbook-held-out records were masked during pathway prediction. NeutroChem Fusion therefore provides a transparent route from exceedance screening to uncertainty-aware source diagnosis, confirmatory-sampling priority and experimentally supported treatment-pathway selection, while independent field validation and site-specific pilot assessment remain necessary. Established method comparisons yielded source-label concordances of 0.8649 for XGBoost-SHAP, 0.6988 for APCS-MLR score mapping and 0.5985 for uncertainty-weighted PMF score mapping. A PHREEQC input-completeness audit found that 0 of 720 records contained the full temperature, major cation, inorganic carbon and redox speciation inputs required for defensible charge balanced simulation.