Accurate large-scale simulations of siliceous zeolites by neural network potentials

The tremendous diversity of zeolite frameworks makes ab initio simulations of their structure, stability, reactivity and, properties virtually impossible. To enable large-scale reactive simulations of zeolites with ab initio quality, we trained neural network potentials (NNP) with the SchNet architecture on a structurally diverse DFT database. This database was iteratively extended by active learning to cover the configuration space from low-density zeolites to high-pressure silica polymorphs including low-energy equilibrium configurations and high-energy transition states. The resulting reactive NNPs model equilibrium structures, vibrational properties, and phase transitions at high temperatures such as thermal zeolite collapse in excellent agreement with both DFT and experiments. The novel NNPs allowed revision of a zeolite database containing more than 330 thousand hypothetical zeolites previously generated employing analytical force fields. NNP structure optimizations revealed more than 20 thousand additional hypothetical frameworks in the thermodynamically accessible range of zeolite synthesis. Additionally, the obtained zeolite database provides vital input for future machine learning studies on the structure, stability, reactivity and properties of hypothetical and existing zeolites.