SE158: Machine-learning prediction of adsorption in nanoporous zeolites

Zeolites are one of the most important catalysts and sorbents in the chemical industry with a wide range of existing and emerging new applications. This talk will present our group’s recent work in developing machine learning approaches to predict adsorption in zeolites. First, I will describe ZeoNet, a robust, transferrable representation based on 3D convolutional neural networks, that can reliably predict the adsorption of complex molecules. Using long-chain hydrocarbons (linear and branched C18 – C30) involved in the hydroisodewaxing process as the target application, ZeoNet demonstrates an excellent performance in predicting Henry’s constants in all-silica zeolites (r2 = 0.977 and MSE = 3.8 in lnkH, corresponding to an error of only 9.3 kJ/mol in adsorption free energy), significantly outperforming other popular representations, including point clouds, vision transformers, and graph-based representations (e.g., M3GNet, MACE).Second, ZeoNet displays an intriguing ability to capture non-obvious structural differences between real and hypothetical zeolites, which offers a potential glimpse into the nebulous question of synthetic feasibility for this class of materials. We apply ZeoNet-based classifiers to ~105 computationally generated structures and demonstrate that these classifiers can achieve a false positive rate of less than 1%, suggesting that intrinsic differences exist despite that they resemble existing zeolites in terms of chemical structure and thermodynamic stability. We hypothesize that the extremely small number of misclassified hypothetical structures are most likely to be eventually synthesizable.Finally, we evaluate several variants from the popular MACE and DeepMD families of machine-learning potentials using grand-canonical and Gibbs-ensemble Monte Carlo simulations. Computed adsorption isotherms and vapor-liquid coexistence curves are compared with experiments and classical forcefield results, revealing their gap in accuracy for predicting these thermodynamic properties.