Machine Learning Potentials for Proton Hopping in Zeolites

Methods

All simulations were performed using a two-stage workflow. First, short ab initio molecular dynamics simulations were carried out using density functional theory (DFT) at the PBE+D3 level of theory with the CP2K code, patched with PLUMED for enhanced sampling via Well-Tempered Metadynamics (WTMetaD). These initial simulations were deliberately short in duration but sufficient to sample chemically diverse configurations and key proton transfer events. From the resulting trajectories, a carefully curated training dataset was assembled using SOAP descriptor-based similarity screening to ensure structural diversity and avoid redundancy. A MLIP was then trained on this dataset using the MACE graph neural network architecture, which learns to reproduce DFT-level energies and atomic forces. To systematically improve the MLIP, an active learning loop was employed: the MLIP was used to run exploratory metadynamics simulations, configurations with high prediction uncertainty were identified via a Query-by-Committee approach, and these structures were re-labeled with DFT and incorporated into the training set. The final MLIP was used for production WTMetaD simulations of up to several nanoseconds, from which converged free energy surfaces were extracted. For studies of aromatic alkylation mechanisms, the accuracy of energy profiles was further elevated to coupled-cluster quality using a correction scheme applied to key stationary points. Long production runs were performed with LAMMP integrated with PLUMED.

Results

This work has produced two key scientific outputs. The first, published in the Journal of Catalysis (2025, vol. 454, article 116658), investigated the protonation equilibrium between ethanol and water at the Bronsted acid site (BAS) of H-ZSM-5 under conditions relevant to biomass conversion. Using MLIP-driven WTMetaD extending to several nanoseconds, we obtained converged free energy surfaces for proton localization as a function of water loading (1-4 water molecules per ethanol). The results reveal a clear, quantitative threshold: with one water molecule, the proton resides predominantly on ethanol; with two, it is shared between ethanol and water; from three water molecules onward, it becomes fully delocalized over the water cluster, mimicking bulk solution behavior. This three-water threshold for hydronium stabilization inside the zeolite pores is a thermodynamically rigorous result that could only be reached through the nanosecond-scale sampling made possible by the MLIP on HPC hardware. The second study, currently submitted, investigated the alkylation of benzene and phenol using cyclohexene as the alkylating agent inside a zeolite framework, focusing on Wheland-type carbocationic intermediates postulated in the mechanism. Combining MLIPs with WTMetaD and CCSD(T)-level energy corrections, we mapped the free energy profiles along the alkylation pathway and found that whether a Wheland intermediate is a true thermodynamic minimum or only a transient species depends critically on the substrate identity, with direct implications for catalyst selectivity in lignin valorization.

Discussion

The original proposal aimed to study bare proton hopping between framework oxygen sites in simple zeolites (Chabazite). During the project, we recognized that this bare proton transfer, while scientifically interesting, requires more sophisticated treatment of nuclear quantum effects and more accurate force fields to produce reliable quantitative results. Instead, we redirected efforts toward catalytically more relevant and more directly impactful questions: how the BAS behaves in the presence of actual reactants under aqueous conditions, and what reactive intermediates govern industrially important reactions. This scientific pivot was productive. The finding that a minimum of three water molecules is needed to stabilize a hydronium ion inside H-ZSM-5 is directly relevant to biomass conversion processes: under reaction conditions, the zeolite pores are rarely fully dehydrated, and knowing the precise hydration threshold that switches the character of the active site from a localized proton donor to a delocalized hydronium reservoir has clear implications for understanding reactivity and selectivity. The Wheland-intermediate study likewise addresses a decades-old mechanistic question with new quantitative rigor: by going beyond static energy calculations and mapping the full free energy landscape at near-coupled-cluster accuracy, we could determine unambiguously which aspects of the alkylation mechanism are kinetically versus thermodynamically controlled. Both studies demonstrate the central value of the HPC resources provided: nanosecond-scale, DFT-quality molecular dynamics simulations of reactive events in zeolites are simply not achievable without large-scale GPU and CPU time, and the results obtained would not have been accessible through any other means.

Additional Project Information

DFG classification: 302-03 Chemical Solid State and Surface Research, Theory and Modelling
Software: PLUMED, LAMMPS, CP2K
Cluster: CLAIX

Publications

Protonation dynamics of confined ethanol–water mixtures in H-ZSM-5 from machine learning-driven metadynamic,
Princy Jarngal, Benjamin A. Jackson, Simuck F. Yuk, Difan Zhang, Mal-Soon Lee, Maria Cristina Menziani, Vassiliki-Alexandra Glezakou, Roger Rousseau, GiovanniMaria Piccini,
https://dx.doi.org/10.1016/j.jcat.2025.116658, February 2026

The Elusive Nature of Aromatic Carbocation Intermediates in Confined Catalytic Environments,
Chintu Das, Princy Jarngal, Fabian Berger, Abhishek Khetan, GiovanniMaria Piccni,
https://dx.doi.org/10.26434/chemrxiv-2026-2krfd/v2, April 2026