Machine Learning Approaches for Heavy Metal Adsorption in Water Treatment Systems: A Systematic Review

Abstract

Introduction: Machine learning (ML) techniques have been increasingly applied to model Heavy metals (HMs) adsorption from aqueous environments; however, variations in datasets, evaluation strategies, and reporting practices have limited the development of generalizable conclusions regarding model performance. This systematic review comparatively evaluated ML applications in heavy metal adsorption modeling and identified factors influencing predictive performance across heterogeneous adsorption systems.

Methods: Peer-reviewed studies published between January 2010 and December 2024 were identified through searches of Scopus, Google Scholar, ScienceDirect, and PubMed. Risk of bias was assessed using the CASP tool, and studies were classified into single-algorithm, ensemble, and hybrid intelligent models.

Results: Of 1,806 records identified, 87 studies met the eligibility criteria. Comparative analysis showed that model performance was largely context-dependent rather than algorithm-specific. Although Artificial neural networks (ANNs), ensemble methods, and hybrid models all demonstrated strong predictive capability, no approach consistently outperformed others across all conditions. Instead, predictive performance was primarily influenced by dataset size, data quality, feature selection, and system complexity. Biochar and activated carbon were the most frequently investigated adsorbents, while arsenic, lead, chromium, and cadmium were the predominant target contaminants.

Conclusions: These findings suggest that model selection should be guided by dataset characteristics and application requirements rather than presumed algorithmic superiority. Standardized reporting, external validation, and greater transparency in model development are needed to improve the reliability and transferability of ML models for real-world water treatment applications