Machine-Learning Driven Ab Initio Enhanced Sampling of Heterogenous Catalytic Reactions

In heterogeneous catalysis, the free energy profiles of reactions govern the mechanisms, rates, and equilibria. Energetics are typically computed using the harmonic approximation (HA), which requires prior identification of critical states. However, applying these methods to chemically reactive systems with complex free energy landscapes, particularly those with large entropy contributions, poses significant challenges. In this work, we describe machine learning approaches, utilizing neural networks and Gaussian Process models, to efficiently sample and directly calculate the free energy surface (FES) of heterogeneous catalytic reactions. We demonstrate these approaches by elucidating the mechanisms of industrially and environmentally significant catalytics reactions on ruthenium: (i) the dissociation of molecular nitrogen for ammonia synthesis and (ii) the hydrogenolysis of dodecane for plastic waste recycling. These examples highlight the role of machine learning in accelerating these calculations to achieve density-functional-theory-level accuracy, while also emphasizing the substantial contributions of configurational entropy of the adsorbate and vibrational entropy of the surface atoms—factors often neglected in HA—to the reaction free energy surface.