Prediction of aluminium distributions in zeolites using basin-hopping and machine learning potentials

In this work, we present a systematic approach to predicting the structures of three sodiated aluminosilicate zeolites (Chabazite, Mordenite and RHO) by combining an accurate neural network potential with a custom basin-hopping algorithm. We show that the low energy, thermally accessible configurations are rare in comparison to the number of possible structures, accounting for a few percent of viable structures. For Chabazite and Mordenite, it is found that there is a tendency for aluminium to preferentially aggregate with 2 or 3 silicon atoms separating aluminium atoms with a strong disincentive to form Al-O-Si-O-Al linkages. However, clear and contrary topological preferences are observed in RHO where Al-O-Si-O-Al linkages are strongly favoured at moderate Al concentrations. Crucially, randomly generated or non-basin hopping approaches to structure generation results in configurations that are highly metastable and activated with respect to minimum energy structures. Using such activated structures for modelling catalysis processes could potentially yield qualitatively incorrect predictions of reactivity and transport properties.