(569ax) Reactive Active Learning for Catalytic Discoveries: Methane Coupling on Titanium Carbide

Ab initio methods for studying catalytic reactions face several limitations. They often require small, simplified surfaces with individual molecules adsorbed at specific active sites, restricting the range and complexity of systems that can be investigated. Ab initio methods for studying heterogeneous catalysis are limited because of the high computational cost and unfavorable scaling with the number of atoms of these methods. These challenges can be addressed with accurate and computationally efficient machine learning potentials. These potentials efficiently map molecular structures onto high-dimensional potential energy surfaces with near-quantum accuracy, significantly reducing computational costs. Unlike empirical forcefields, they utilize molecular descriptors, eliminating the need for explicit bonding terms. However, training these potentials to encompass various chemical reactions is often extremely inefficient. We propose a method addressing this by automatically generating potential reactions and employing transition-state (TS) finding techniques without explicitly specifying reaction pathways. We employed the DeePMD formalism to train deep-learning potentials in an iterative active learning scheme. These deep-learning potentials were used to explore reactive space via automatic generation of many different possible reactions and TS finding codes like the nudged elastic band method. This active learning scheme was tested for the methane coupling reaction on various Ti_xC_y surfaces.