An Ensemble Learning Framework Utilizing Fusion Molecular Fingerprints for Pollutant Removal Prediction in Advanced Treatments

Accurate prediction of pollutant degradation kinetics is essential for assessing and optimizing water treatment processes. However, conventional quantitative structure–activity relationship (QSAR) models are often limited by incomplete chemical representations derived from single-molecular fingerprints. Here, we developed a fingerprint-fusion ensemble learning framework to accurately predict pollutant reactivity in advanced treatments, using ozone oxidation and zero-valent iron (ZVI) reduction as case studies. Our framework decoupled the prediction task by employing specialized base learners to extract intrinsic structural reactivity from complementary molecular fingerprints, capturing features from the composition to conformation. A meta-learner subsequently integrated these structural predictions with the environmental variables. The ensemble framework demonstrated superior predictive accuracy, achieving test R 2 values of 0.96 for ozonation and 0.80 for ZVI reduction, outperforming the best single-fingerprint baseline model by more than 5%. Multidimensional interpretation analysis further elucidated the underlying prediction logic, identifying key compositional, topological, and conformational drivers of pollutant reactivity. Finally, the framework was deployed as an interactive Web platform, providing an accessible tool for reactivity prediction and mechanistic exploration. This work establishes fingerprint-fusion ensemble learning as an effective strategy for predicting and interpreting pollutant reactivity in advanced water treatments.