Atomistic Landscape of Pt Nanoparticles via Machine Learning: How Size Effect and Hydrogen Adsorption Govern Structural Ensembles and Catalytic Activity

Understanding the atomistic structure and fluxionality of Pt nanoparticles under reactive conditions is essential for rational design of effective catalysts, yet their structural complexity presents a great challenge. In this work, we combine grand canonical global optimization methods and machine learning potentials to explore the atomistic landscape of nanometer-sized (1∼2 nm) Pt nanoparticles under a pressure of hydrogen, resulting in a comprehensive library of Pt_xH_y nanoparticles with over one million low-energy metastable structures. We found that hydrogen adsorption drives a size-dependent transformation from an amorphous to a crystalline structure, leading to sharp phase transitions for smaller nanoparticles and smooth transformations for larger ones. This behavior is governed by a competition between distinct core configurations, as well as the formation of rigid and fluxional local domains, where stability is dictated by specific surface motifs at low H coverage and by the crystalline core at high H coverage. By applying this structural library for reactivity modeling of methane dehydrogenation and ethylene hydrogenation, we show a marked discrepancy between the abundance of a surface site and its catalytic contribution, indicating that the active sites are rare, structurally distinct motifs, not the most common sites.