Exploring the Reaction Network of Acetic Acid in Supercritical Water via Machine Learning Interatomic Potential

Supercritical water oxidation offers promising solutions for waste treatment, but understanding its complex molecular reaction mechanisms remains challenging due to extreme experimental conditions. We compare two computational approaches, a machine learning potential (NequIP) and a reactive force field (ReaxFF), to model acetic acid oxidation in supercritical water, a key industrial process. While ReaxFF predicts the apparent activation barrier closer to experimental measurements, NequIP more accurately reproduces the observed product distributions and reaction pathways. NequIP successfully captures the experimentally confirmed radical reaction mechanisms and complete oxidation behavior, whereas ReaxFF overestimates intermediate stability and favors incomplete oxidation. Both models correctly predict enhanced reaction rates when hydrogen peroxide replaces oxygen as the oxidant though with different effects on specific reaction steps. These findings demonstrate that machine learning potentials can effectively combine quantum mechanical accuracy with computational efficiency for modeling complex reaction networks, providing valuable insights for optimizing industrial oxidation processes despite current limitations in predicting absolute energy barriers.