Application of machine learning methods for phenols degradation by Fe-based catalytic Fenton-like process.

There are complex interactions among the catalyst composition, material structure, aqueous matrix, and operational conditions in the Fe-based catalytic Fenton-like system for phenols removal, which limit the accurate prediction of removal performance and the explanation of mechanisms across different reaction systems. Herein, based on the literature, a comprehensive dataset comprising information on phenols properties, the composition and structure of Fe-based catalysts, aquatic matrix, and reaction conditions was constructed to develop prediction models for phenols removal efficiency under various generalization scenarios. Through grouped nested cross validation, the predictive performance of various machine learning and deep learning models was firstly compared. Results showed model performance was clearly dependent on the validation group: Long Short-Term Memory achieved the highest pooled out-of-fold R2 in the curve-group (0.883) validation; Random Forest in the system-group (0.708) and catalyst-group (0.681) validations; and Extreme Gradient Boosting (0.589) in the paper-group validation. Furthermore, integrated SHAP, PFI, and VIF analysis revealed that reaction time and pH were the most stable model-related factors. The reaction time primarily reflected the cumulative removal process resulting from the continuous conversion of phenols and intermediate products, while pH might further modulate the H2O2 activation efficiency by affecting the Fe(II)/Fe(III) cycle, Fe hydrolysis precipitation, surface passivation, and the availability of active Fe sites. The non-monotonic response of Fe-based dose further indicated that an increase in Fe-based surface sites might not necessarily lead to corresponding improvement in removal efficiency, and its effect might be jointly regulated by pH, phenol and H2O2 concentrations, and interfacial mass transfer. Finally, this data-driven framework effectively overcomes the limitations of empirical models, providing reliable technical support for the accurate prediction of Fe-based nanocomposites performance in the Fenton-like system under complex operation conditions.