A Decade of Machine Learning for Adsorption of Emerging Contaminants: A Systematic Review of Algorithms, Performance, and Methodological Transparency

Machine learning (ML) is increasingly applied to predict and optimize adsorption-based removal of emerging contaminants (ECs) from water, yet the literature remains fragmented across diverse algorithms, adsorbents, and pollutant classes, and existing reviews of the topic report neither a quantitative synthesis of performance nor a cross-study interpretability analysis. This systematic review, conducted following PRISMA 2020 guidelines, screened 2,164 records from Scopus and PubMed and analyzed 202 peer-reviewed studies published between 2015 and 2025. Among 59 unique algorithms, artificial neural networks (ANN) dominated (64.9%), though ensemble methods (particularly random forest and XGBoost) are rapidly gaining adoption. Inverse-variance weighted meta-analysis of 110 studies revealed pooled R2 values ranging from 0.998 (ANFIS) to 0.929 (CatBoost), with extremely high heterogeneity ( 98\%]]>I2>98%) precluding universal algorithm recommendations; funnel plot analysis revealed patterns consistent with publication bias. A cross-study synthesis of SHAP-based feature importance from 15 explainable ML studies identified initial concentration, BET surface area, and pH as universally important predictors, consistent with classical adsorption theory, while revealing system-specific drivers (elemental ratios for biochar, molecular weight for PFAS) that extend beyond traditional frameworks. Antibiotics, pharmaceuticals, NSAIDs and analgesics together accounted for 81.7% of target contaminants, while PFAS (9.4%) and microplastics (5.9%) remained substantially underrepresented. Critical methodological concerns persist: among studies reporting a dataset size, 40% used fewer than 50 data points, and an audit of 30 randomly selected papers read in full found that although 90% describe a validation scheme, every one is an internal partition of a single experimental dataset and none validates against independently collected data. To address these gaps, we propose a 22-item Minimum Reporting Checklist for ML in Adsorption Studies (MRCAS), a technology readiness assessment for six frontier architectures (PINNs, GNNs, transfer learning), and five evidence-based priority directions to advance the field from prediction-focused modeling toward practical decision-support tools for water treatment.