Enhanced sampling molecular dynamics simulations of biomass alcohols conversion over zeolites

Methods

Integrating the state-of-the-art variant of metadynamics, known as Well-Tempered Metadynamics, with deep neural network-based machine learning offers significant computational speedups, reducing the load of DFT-level calculations on CPUs. To achieve AIMD-level accuracy, we will refine the deep potential using an active learning strategy. This method enables the construction of reactive potential energy surfaces of ab initio quality, enhancing sampling and extending the scope of AIMD simulations. The approach involves iteratively extending the dataset and refining model potentials through training, exploration, and labeling. Initially, an ensemble of deep potentials is trained on a small set of structures from a DFT reactive trajectory. Following this, multiple DP-based MetaD simulations are conducted, and subsets of new configurations are selected for labeling. These labeled configurations are added to the dataset, iteratively extending and refining the model.

Results

We generated training datasets for four zeolite topologies—Faujasite, Chabazite, Beta, ZSM-5, and Gismondine—focusing on ethanol from lignocellulosic biomass. Using these datasets, we developed a reactive DP potential to obtain the free energy surface for cyclohexanol dehydration in H-BEA and H-MFI zeolites with 13 water molecules, concluding an E1 reaction mechanism in both zeolites. To refine this potential for transferable dehydration reactions, extensive metadynamics simulations requiring significant computational time are necessary. A similar work in the spirit of this research was conducted which helps to shed light on the dehydration of cyclohexanol in apolar solvent, decalin. It was revealed that dehydration rates are 2-3 times higher in decalin than in water for BEA and FAU, whereas MFI showed similar reactivity in both solvents. We generated DP models for comprehensive sampling of cyclohexanol and decalin isomers (cis and trans) in H-MFI zeolites, also characterizing single-component adsorption at 0 K to determine energies and geometries. Unsupervised machine learning techniques were applied to analyze the local environment around Brønsted acidic sites using SOAP descriptors, providing a molecularlevel understanding of adsorption mechanisms. These calculations are crucial for understanding the impact of micropore environments on sorption and catalysis in both polar and apolar solvents, such as water and decalin.

Discussion

The results provided us with an unprecedented molecular-level understanding of the reactive process of alcohol dehydration in the pores of acidic zeolites, which aligns with experimental thermodynamic and kinetic measurements. Moreover, they allow us to delve into the intricate mechanistic aspects involved in the reactivity, such as the influence of water molecules or organic solvents present in the pores and the confining effects of the zeolite’s walls. These walls exhibit dual hydrophobic interactions: repulsive interactions with the water molecules and stabilizing dispersion interactions with the organic substrates. The reaction mechanism confirms the mechanistic hypothesis that holds for the same reaction in solution but shows a direct influence of the pore confinement effects. These effects can alter the height of the activation barrier by locally stabilizing the activated complex, similar to biological enzymes. This aspect is fascinating and poses a fundamental challenge in the basic understanding of these processes, necessitating further investigation with more simulations extending the study to a larger number of zeolite topologies. Once these fundamental aspects are clarified, we will move on to the next step concerning the use of the dehydration products for the alkylation step, thus closing the circle for studying the production of heavier molecules to be used as commodity chemicals or blended into biofuels. This research demonstrates great potential, providing both a fundamental scientific understanding of chemistry and catalysis theory, and an atomistic-level understanding of real-life catalysts. It also offers insights into the influence of the material on the process and its kinetics, a critical aspect at the practical and industrial levels.

Additional Project Information

DFG classification: 303 Physical and Theoretical Chemistry
Software: LAMMPS, PLUMED, DeePMD-kit, CP2K, Dscribe
Cluster: CLAIX

Publications

Wenda Hu, Nicholas R. Jaegers, Princy Jarngal, Sungmin Kim, Feng Chen, Oliver Y. Gutierrez Tinoco, Ruixue Zhao, Yue Liu, Qiang Liu, Donald M. Camaioni, Yong Wang, Jian Zhi Hu, GiovanniMaria Piccini, and Johannes A. Lercher, Adsorption and Desorption of Decalin, Water, and Cyclohexanol in Zeolites by In Situ ²H MAS NMR and Molecular Modeling with Machine Learning, in preparation

Thesis
Princy Jarnagal,
Combininng molecular simulations and machine learning to understand catalytic processes in zeolites