Accurate prediction of relative free energies and reaction rates in zeolite catalysis is highly challenging due to both the complexity of the interatomic interactions as well as the necessity of explicit high-temperature configurational sampling.
Including both energetic and entropic effects in atomistic simulations is up to now impossible because the quantum mechanical calculations required for an accurate estimate of the energetics are too expensive to repeat millions of times, as necessary during high-temperature sampling. Here, the project proposes to construct the first ever reaction free energy profiles which approach chemical accuracy using a combination of highly accurate quantum mechanical calculations, enhanced sampling molecular dynamics, and state of the art machine learning (ML) potentials.
Parameterisation of the potential is performed in an efficient manner using transfer learning from a relatively inexpensive density functional theory (DFT) reference towards the highly accurate random phase approximation (RPA) level of theory, after which it is used in long umbrella sampling simulations in order to estimate the relative free energy differences and reaction rate.
The computational efficiency allows us to compare the reaction free energies of a topical hydrocarbon reaction across hundreds of frameworks and aluminum sites in order to understand which configuration maximizes the desired intermediate formation.
Ghent University, Belgium.