We present an emerging workflow that leverages tuned machine-learned interatomic potentials (MLPs) to accelerate reaction discovery in heterogeneous catalysis. Traditional first-principles methods are often too computationally intensive for exploring complex reaction networks on catalyst surfaces. By training system-specific MLPs—guided by quantum chemistry and active learning—we efficiently sample relevant intermediates and transition states at scale. These reactive pathways are then linked to mesoscopic kinetic models to assess their impact on catalyst activity, selectivity, and stability under realistic conditions. We highlight how this approach enables mechanistic insight and supports multiscale simulations for materials screening and process optimization. The workflow, currently under development in the Amsterdam Modeling Suite, lays the foundation for integrating data-driven exploration with physics-based reaction modeling, paving the way for predictive catalyst design.
