Energy optimization in electrochemical oxidation for wastewater treatment using interpretable machine learning

• Built a 255-sample EO dataset from literature and laboratory sources. • Trained and compared RF, SVR, and XGBoost models for EO energy prediction. • XGBoost achieved strong accuracy (R² = 0.97; RMSE = 25.7 kWh·m⁻³). • Current density and pollutant class were identified as dominant drivers. Electrochemical oxidation (EO) promises broad-spectrum removal of recalcitrant pollutants from wastewater but is often penalized by high energy demand and uncertain scale-up. This study presents a data-driven framework that links EO operating conditions to energy consumption and removal performance using a curated dataset of 255 experimental conditions collected from the literature and laboratory reports. Three regression models, Random Forest, SVR and XGBoost were trained, compared, and tuned through nested cross-validation, with XGBoost yielding the best predictive accuracy (5-fold CV R² = 0.9727, RMSE = 25.7 kWh·m⁻³, MAE = 8.95 kWh·m⁻³). Interpretability analyses (feature gain, SHAP summary, PDPs and ablation tests) show that current density and pollutant class jointly explain the majority of variance, with current density alone accounting for the largest single share. A leave-one-study-out experiment reveals that prediction degrades when extrapolating to pollutant categories not represented in the training set, highlighting a clear limitation for broad extrapolation. The complete dataset, trained models and reproducible code are provided in the supplementary repository, supporting transport and data-driven optimization of EO systems for wastewater treatment.