Reactive Machine-Learned Potentials Improve Predictions of Water Diffusion Behaviour in Defective Industrial Zeolite HY
Year: 2026
Type: preprint
Abstract: Dealuminated HY (USY) zeolites are of major industrial catalytic importance, and the defects they contain, such as extraframework aluminum (EFAL) and silanol groups, strongly influence the behaviour of adsorbed species. Here, we show that reactive neural network-based machine learned interatomic potentials (MLIPs) provide a substantial improvement in the simulation of water diffusion in these catalysts. They yield markedly improved agreement with quasielastic neutron scattering (QENS) measurements of mobility for models of intact framework regions, refined further once EFAL and silanol defects are included explicitly, capturing the additional hindrance these sites impose on water motion. The ability of MLIPs to generate and stabilise realistic defect structures constitutes a step change in the molecular level and nanoscale modelling of industrial zeolites. This combined capability enables the quantitative investigation of defect-specific effects on nanoscale diffusion and provides a new approach for accurately modelling transport in industrially relevant zeolite catalysts.
Source: ChemRxiv
Institutions: Charles University, University of Bath, Institut Laue-Langevin
Language: English
Subfield: Inorganic Chemistry
Field: Chemistry
Domain: Physical Sciences
Open Access status: green
Awards: G1519889414, G4012625075, G5960638817
OpenAlex ID: W7223324583