A Brønsted-Evans-Polanyi Relationship-Guided Machine Learning Model for Predicting H2 Activation Mechanism over...

H2 activation mechanism governs the activity, selectivity, and stability of CO2 hydrogenation over In2O3 surfaces, which varies with the coverages of oxygen vacancies under the realistic reaction conditions. Herein, we systematically map H adsorption and H2 dissociation on cubic In2O3(111) across a broad range of oxygen vacancy coverages, explicitly distinguishing heterolytic cleavage at In-O pairs from homolytic cleavage at In-In pairs. Our results demonstrate that these two mechanisms follow distinct kinetic regimes rather than a universal Brønsted-Evans-Polanyi (BEP) relationship. The BEP slopes reveal that the descriptors of the initial state are more informative for heterolytic barriers, while those of final state better capture homolytic barriers. This mechanistic insight effectively narrows the candidate descriptor pool for predicting activation barriers across various surface states. SISSO and benchmark machine learning models consistently support this state dependent descriptor preference, highlighting the underlying chemical necessity of matching descriptors to TS character. Our results also reveal that H2 activation on catalytically relevant pristine and moderately reduced surfaces is mainly achieved through heterolytic cleavage at In-O sites, while the homolytic cleavage at In-In sites is promoted and becomes prominent under deep reduction. However, excessive reduction does not monotonically promote H2 dissociation. This work provides a physically interpretable protocol for mechanism-aware descriptor screening to elucidate hydrogen activation on metal oxides, guided by mechanistically resolved BEP relationships.