(159c) Machine-Learning Driven Exploration of Catalytic Reaction Network

The adsorption energy and reaction barrier are two primary inputs for microkinetic modeling of catatlytic processes. However, when involved reaction intermediates are relatively large and diverse, e.g., in syngas conversion, different types of complexity are associated with attaining these parameters. For medium sized molecule, a multitude of possible binding motif often leads to complex potential energy surfaces (PES) with many competing adsorption conformation. This requires the use of global optimization algorithms, which entails computational costs that are prohibitive at the DFT level. Similarly, sophisticated free energy barrier simulations using, e.g., umbrella sampling, are often replaced by the less rigorous harmonic approximation for computational reasons. To tackle these issues, we present a protocol that generates surrogate machine learning potentials for surface chemistry on-the-fly. This is applied to both global optimization and free energy calculations pursuing minimal human intervention and computational cost by iteratively updating the training set with the configurations explored by the algorithm.