Computing accurate rates for reactions occurring in complex environments is very challenging, primarily due to the exponential dependence of kinetic constants on activation enthalpy and entropy. Chemically accurate enthalpies require advanced electronic structure methods, whereas configurational freedom must be taken into account for entropies using dynamic methods such as transition interface sampling (TIS). Combining both was thus far virtually impossible for the prohibitive computational cost. Yet, computing accurate rates is essential not only to explain experimental outcomes, but also to guide and predict them. This project will harness the power of state-of-the-art machine learning potentials (MLPs) to model activated events with chemical accuracy. We focus on zeolite catalysis, but the methods are of general applicability for activated events in complex environments. MLPs will allow us for the first time to combine advanced simulation methods such as TIS and grand canonical Monte Carlo to derive chemically accurate reaction models, first for simple molecular systems and then in well-defined zeolite environments. Finally, all methodological advances will enable to tackle the industrially relevant zeolite-catalyzed direct alcohol amination, conducted under challenging high-loading conditions. The project will be executed in synergy with leading theoretical and experimental partners to validate our models and provide additional insights into the investigated reactions.