Machine Learning Unlocks the HOMO Energy Level as the Master Regulator of Micropollutant Decontamination in Covalent Organic Frameworks-Based Photocatalysis.

Accurately predicting micropollutant degradation kinetics during water purification via heterogeneous photocatalysis remains a major challenge. Here, the highest occupied molecular orbital energy (EHOMO) of typical micropollutant is established as the dominant descriptor in the TpBpy-COF system governing photocatalytic oxidation rates, using a versatile and robust TpBpy-COF (β-ketoenamine covalent organic frameworks) platform that achieves efficient phenol removal (100% within 15 min, 85.4% mineralization) with excellent environmental adaptability. Employing this model system and 15 structurally diverse pollutants (10 phenolic and 5 nonphenolic compounds), it is demonstrated that while the catalyst generates identical reactive species (•O2- and h+), their utilization efficiency is dictated exclusively by EHOMO via a dual-pathway mechanism: surface-confined hole oxidation governed by micropollutant adsorption affinity, and electron donation to superoxide radical that triggers cascade hydroxyl radical generation. Integration of 12 descriptors and machine learning-derived quantitative structure-activity relationships (QSAR) model, reveals that EHOMO and ionization potential (IP) are highly correlated with the degradation rate constant (R2 > 0.86). Further, the Extremely Randomized Trees (ET) algorithm performs best and SHapley Additive exPlanations (SHAP) analysis identify EHOMO as the overwhelmingly dominant predictive feature. This work establishes a predictive framework linking micropollutant electronic structure to degradation kinetics, shifting the paradigm from a catalyst-centric to a catalyst-pollutant electronic coupling view.