Automated machine learning and SHAP-based interpretation of PFOA removal via electrochemical oxidation
Haitham Elnakar
Abstract
The electrochemical oxidation of perfluorooctanoic acid (PFOA), a persistent and toxic environmental pollutant, presents significant operational complexities due to nonlinear interactions among numerous process variables. This study introduces a fully automated, robust machine learning (ML) framework based on Fast Lightweight AutoML (FLAML), integrated with SHapley Additive exPlanations (SHAP) to rigorously optimize and interpret PFOA electrochemical oxidation performance. The FLAML-optimized XGBoost model demonstrated exceptional predictive accuracy (RMSE=3.97, R 2 =0.98), significantly outperforming systematically tuned traditional ML models such as Random Forest, Gradient Boosting, and Deep Learning architectures. Rigorous statistical validation confirmed the superior generalizability and stability of the FLAML-optimized model, while SHAP-based interpretability analyses revealed electrolysis time and anode material as primary drivers of treatment efficacy, with electrolyte concentration and current density forming the next tier; consistent with established electrochemical degradation kinetics. Notably, FLAML-optimized XGBoost achieved a 72 % reduction in computational overhead compared to manually tuned models, establishing a reproducible, interpretable, and computationally efficient benchmark in environmental ML applications. The framework provides a systematic approach that improves predictive accuracy, enhances interpretability, and strengthens practical applicability to support optimization of remediation technologies and environmental decision-making.