Reactive Neural Network Potential for Aluminosilicate Zeolites and Water: Quantifying the effect of Si/Al ratio on proton solvation and water diffusion in H-FAU

Abstract

Acidic zeolites are one of the most important catalysts. In many of their catalytic applications, the mode of interaction with water heavily influences their activity, efficiency, and durability as a catalyst. Despite the recent (first principles) computational efforts to understand the mechanistic underpinning of the water-zeolite interactions, it is still prohibitively expensive to carry out comprehensive studies employing realistic zeolitic models. Therefore, we developed a reactive neural network-based potential for aluminosilicate zeolites in the protonic form including their interaction with the aqueous solution that has a capacity to accelerate simulation by orders of magnitude while retaining the reference level of accuracy. To showcase its potential, we used it to determine how multiple factors (aluminum content, water loading and temperature) influence the proton solvation and water dynamics in one of the industrially most important acidic zeolites, the faujasite (FAU). We found that Si/Al ratio is a significant determinant of the water diffusivity, water capacity to solvate protons in the nanopores, and the zeolite hydrolytic stability. We expect that many of these findings are readily extendable to other acidic zeolites in interaction with water.

Probing the Role of Acid Site Distribution on the Water Structure in Aluminosilicate Zeolites: Insights from Molecular Dynamics

“…As the water loading increases, the average distance in all configurations decreases, indicating protonation of adsorbed water and solvation into the cluster. This observation aligns with previous literature at high Si/Al ratios, 29 and in faujasite, 45 which reported similar trends of proton solvation assisted by coadsorbed water. We noticed that all Al distributions exhibit a similar trend in proton distance with near-saturation loadings having the average proton locations closest to the pore center.…”

“…Roy et al developed reactive potentials for silicates using an equivariant neural network that captured water deprotonation and silica dimerization reactions. 43 Neural network potentials for aluminosilicate zeolites have been developed and applied by Erlebach et al and Saha et al, 44,45 which included interactions between water molecules and Brønsted acid sites using 10 different zeolite structures. By understanding the molecular interactions between the solvent and zeolites, simulations can open new avenues for optimizing reaction kinetics to enable the development of more efficient and selective catalysts.…”

The need for operando modelling of 27Al NMR in zeolites: the effect of temperature, topology and water

“…Nevertheless, the training and testing of MLPs is not straightforward and issues of transferability of MLPs between zeolites of various compositions, topologies, counterions and water content remain an active area of development. 65 It is likely that conclusions drawn here regarding the necessity of operando modelling for 27 Al NMR spectra can be generalized also for other zeolite extra-framework and framework elements (Si and O in particular) and even for different solids. The importance of operando modeling for zeolites is enhanced by the fact that zeolites are microporous materials and they contain water in their channels.…”

“…If longer runs or if too many Al congurations are required for a particular study, MD simulations can be possibly performed with reliable machine learning potentials (MLPs), as recently demonstrated, e.g., for siliceous and aluminosilicate zeolites. 65,67,68 This will alleviate the sampling problem which currently hinders dynamical simulations, by extending the available timescales by several orders of magnitude. Nevertheless, the training and testing of MLPs is not straightforward and issues of transferability of MLPs between zeolites of various compositions, topologies, counterions and water content remain an active area of development.…”

Constructing Collective Variables Using Invariant Learned Representations

“…Obviously, the generation of the NNP comes with the up-front cost of generation of additional reference (DFT) data. However, a recent work suggests that low tens of thousands of reference data points should be sufficient to train a robust NNP model for a specific reaction, while our own recent study shows that a general NNP covering broader range of chemical compositions and (reactive) transformations, such as the NNP for aluminosilicate zeolites in interaction with water used herein, needs approximately 10 5 –10 6 single-point DFT calculations. However, this up-front cost is expected to be amortized or even overcompensated by (i) the superior performance of the CVs based on pretrained representations, which allow for more efficient acceleration of the biased simulations, e.g., by speeding up the convergence of CVs in iterative refinement schemes, , and (ii) the replacement/acceleration of reference (DFT) calculations by orders of magnitude faster NNP calculations generating new structural data points for the autoencoder to refine the CVs.…”