Tailored machine learning potentials for the in silico design of zeolites
Zeolites are well known microporous aluminosilicates produced industrially at the Megaton scale owing to their large number of applications ranging from catalysis through adsorption and separation to ion-exchangers. Optimization of existing and discovery of novel zeolites requires a detailed, atomic-level understanding of the zeolite stability under high temperatures and (water vapor) pressures [1-3]. However, the targeted, in silico optimization of the zeolite synthesis is hampered due to the high computational costs of ab initio simulations and the limited accuracy of available empirical force fields.
Therefore, this work focuses on the training of deep tensor neural network potentials  (NNP) for accurate modeling of the potential energy surface of siliceous zeolites. For this, the training procedure employs a comprehensive dataset containing structures, energies, and forces calculated at the density functional theory (DFT) level for several zeolite frameworks, bulk silica polymorphs, silica glass, and surface slabs. For all systems, sampling of the configuration space includes both close to equilibria and non-equilibrium structures. The resulting NNPs accurately reproduce the energies and vibrational properties of various silica structures in good agreement with DFT simulations. Subsequently, structure optimizations at the NNP level are applied to more than 330k (hypothetical) frameworks  yielding lattice energies with unprecedented accuracy and providing vital information for the assessment of zeolite stability and durability . In addition, the new NNPs facilitate rapid calculations of (free) energies of zeolites under high temperatures and pressures, including phase transitions such as zeolite amorphization .
Extension of the obtained NNP for consideration of zeolite-water interactions and heteroatoms such as Ge and Al is expected to enable accurate and realistic modeling of the thermodynamic stability of zeolites under synthesis and operating conditions.
|Category||Short file description||File description||File Size|
|Extended Abstract||Abstract pdf||PDF version of the abstract including a figure||225 KB||Download|