(192e) Machine Learning-Directed Advanced Sampling Simulations of Reactions in Condensed Phases

Reactions in condensed phases are at the heart of many physical and chemical processes in molecular synthesis, heterogeneous catalysis, and material processing. The mechanism, rates, and chemical equilibria can be understood from free energy profiles of reactions. Predicting the energetics has involved a trade-off in accuracy between the enthalpic (i.e., interatomic potential) and the entropic (i.e., degrees of freedom) contributions to the free-energy. Traditionally, computational efforts in molecular reactions have been largely focused on the former. The latter, however, becomes increasingly important as the environmental effect and the size of the reactive species grow.

Here, we introduce two advanced sampling approaches that combine machine-learning algorithms and electronic structure calculations, to bridge the gap between accuracy and system size for reactions in condensed phases. First, we present an adaptive-biasing technique using neural networks, CFF-AIMD, for computing free energy surface with DFT-MD accuracy. We demonstrate how the neural network sampling method allows efficient sampling and direct computation of free energies through examples in molecular reactions on metallic surfaces and in liquid phases. Second, we develop a hierarchical resolution scheme for force field optimization with advanced sampling simulations. The method is applied to probe the formation, migration, and annihilation mechanisms of spin defects in wide-band gap semiconductors, which are relevant for understanding and designing material platforms for quantum technologies.