G-4508

Evaluation of Machine Learning Algorithms for Predicting and Modeling Arsenic Adsorption from Water by Various Adsorbents: a systematic review

 Roghayeh Hatami ¹ ℗, Omid Yousefianzadeh ² ©, Arash Dalvand ¹   

 
Abstract

2. Introduction: The efficiency of arsenic adsorption is strongly influenced by factors such as pH and adsorbent properties, making accurate modeling challenging. Recently, machine learning (ML) has emerged as a powerful tool for capturing these complex interactions with high predictive accuracy. This review examines ML applications in arsenic removal, highlighting their advantages, limitations, and the need for standardized methods to ensure reliability and comparability across studies. 3. Search Strategy: The review adhered to PRISMA guidelines, involving a comprehensive literature search across PubMed, Scopus, Google Scholar, and Science Direct, covering publications from 2010 to December 2024. The search terms included “machine learning,” “Arsenic,” “adsorption,” “water treatment,” and “heavy metals.” Studies were included if they were published in English, employed ML models to predict or model Pb adsorption processes, and reported performance metrics such as R² or RMSE. Publications were excluded if they were non-English or did not utilize ML in modeling or prediction. Data extraction was carried out using a researcher-developed checklist. The risk of bias and study quality were evaluated using the CASP tool, with only studies scoring 7 or higher included in the review. 4. Results: A comprehensive search yielded 1142 articles, of which 12 studies were included following rigorous screening and quality assessment. ML models, especially XGBoost, CatBoost, Random Forest, and hybrid ANN architectures, have proven highly effective in predicting arsenic and heavy metal adsorption on biochar and engineered adsorbents. XGBoost consistently achieves superior performance, with R² values reaching up to 0.99 and RMSE as low as 0.43–11.78, depending on the system. In multiple studies, XGBoost outperformed other algorithms in modeling As(III), As(V), and multi-metal adsorption. Hybrid models like SVM-ANN (R² = 0.987) and ensemble methods such as GBDT (R² 0.99) further demonstrate high predictive accuracy. These results underscore the robustness of ML, particularly gradient boosting and integrated models, in optimizing and simulating adsorption processes with minimal error. 5. Conclusion and Discussion: The high predictive accuracy of XGBoost (R² up to 0.99) and ensemble models highlights their reliability in simulating complex adsorption processes. Minimal errors (e.g., RMSE: 0.43–1.24) confirm the strong agreement between predicted and experimental values. These results suggest that ML, particularly gradient boosting and hybrid models, can effectively replace traditional modeling approaches in adsorption studies.

Keywords: machine learning; heavy metals; adsorption; water treatment; Arsenic