Harnessing interpretable machine learning: SHapley additive exPlanations (SHAP)-driven insights, transformative impact, and controversies in adsorption-based environmental remediation

Abstract

Machine learning (ML) provides powerful predictive capabilities for environmental remediation, enabling the diagnosis of contamination sources and optimization of treatment processes for pollutants such as heavy metals, dyes, and pharmaceuticals. However, the black-box nature of many ML models limits their mechanistic interpretability, hindering application in process design. This review systematically synthesizes and critically evaluates the use of Shapley Additive exPlanations (SHAP) to address this gap in adsorption-based water treatment. Whereas previous reviews have established the broad utility of ML, a dedicated assessment of SHAP's methodological aspects and its role in deriving mechanistic insight is lacking. The consolidated evidence from diverse studies shows that SHAP analysis reliably identifies key predictors of adsorption behavior, including parameters such as surface area and pH that determine contaminant–adsorbent interactions. A critical review of studies addressing controversies and divergent perspectives in SHAP-based interpretability revealed that, although SHAP is widely employed to extract mechanistic insights, its application frequently overlooks important methodological limitations. The review concludes by outlining future research directions for leveraging SHAP to advance fundamental understanding and optimize remediation strategies.