Machine learning-guided inverse design of persulfate catalysts: From global screening to targeted single-atom...

Achieving both high activity and precise control over reaction pathways (radical vs. non-radical) in persulfate-based advanced oxidation processes (PS-AOPs) remains a key challenge. Here, we report a machine learning (ML)-guided inverse design framework that bridges computational screening and experimental validation for scenario-specific catalyst development. To this end, we developed two complementary global models: a Global Mechanism Predictor (GMP) classifier with 80% accuracy, and a Global Activity Predictor (GAP) regressor with a test set R² of 0.808. These models collectively identified iron-based single-atom catalysts (Fe-SACs) as the optimal platform for non-radical oxidation. To validate this discovery and guide targeted synthesis, we constructed a specialized Fe‑Single‑Atom Activity Predictor (Fe‑SAAP) model, achieving superior predictive performance (test set R² = 0.908). Interpretable ML analysis identified the Fe-N coordination number as the key structural descriptor. Leveraging this insight, virtual screening locked its optimal range at 3.70-5.24. Guided by these targets, we synthesized a series of Fe-N-C catalysts with different Fe loadings and pyrolysis temperatures. The candidate Fe‑N‑C‑1‑1000 (Fe(acac)₃ dosage of 1.2 mM and pyrolysis temperature of 1000 °C) matched the predicted structural characteristics and exhibited an experimental degradation rate (0.34 min^-1) in excellent agreement with the model's prediction (0.353 ± 0.05 min^-1). The catalyst also demonstrated excellent pH adaptability and strong interference resistance. This successful case study completes the inverse design cycle for Fe-SACs and establishes a transferable framework that bridges data-driven discovery with targeted synthesis for scenario-specific environmental catalysts.