Exploring the Intricacies of Glycerol Hydrodeoxygenation on Copper Surface: A Comprehensive Investigation with the Aid of Machine Learning Force Field

Abstract

The utilization of biomass for feedstock chemicals often relies on transforming hydroxyl-containing molecules. One such example is glycerol, which can undergo a selective hydrodeoxygenation reaction to produce propanediol, a valuable chemical precursor. Hence, glycerol’s hydrodeoxygenation reaction combines immediate industrial application with the foundation of fundamental research into the reaction class relevant for sustainable feedstock. Given the complex nature of large organic molecules, most modeling work in heterogeneous catalysis focuses on the reactivity of small (C1–2) organics exclusively. Glycerol, characterized by its C3-backbone, has 75 distinct gas-phase conformers with energy variation up to 0.1 eV (Callam, C. S. J. Am. Chem. Soc. 2001, 123, 11743–11754). When considering its 11 reactive bonds (C–O, C–H, and O–H), the modeling of glycerol’s reactivity spans an extensive conformational and reactive space. High computational costs of density functional theory simulations restrict exhaustive exploration of the factorial reaction space, leading to limited insights into the hydrodeoxygenation (HDO) mechanism and hindering rational catalyst design. Therefore, to date, there is no systematic study focusing on comprehensively sampling the energetics of surface conformers of glycerol and their reactivity. In this study, we employ a message-passing graph neural network architecture (MACE) to develop a machine-learned force-field (MLFF) potential, incorporating active learning to investigate the impact of conformational complexity on the reaction network of glycerol HDO on a Cu(111) surface. Following five iterations, our trained MLFF model accurately predicts surface bound structures with a root-mean-square accuracy of 0.04 eV (<0.6 meV/atom total energy), essential to accurately determine conformational minima of 24 metastable and 26 intermediate states along seven competitive pathways. Conformational sampling uncovers the intricate nature of the complex energy landscape, where conformers with multiple shallow minima lead to nontrivial trends in the transition state energies connecting them. Notably, the investigations predict lower activation barriers for O–H bond scissions of glycerol structures with α- and γ-backbone as compared to β-backbone. This is significant in the case of scission of secondary O–H glycerol bonds where the activation barrier varies up to 0.44 eV depending upon the initial glycerol structure motif. Altogether, we identify dehydrogenation–dehydration–hydrogenation as the dominant pathway resulting in PDO formation on the Cu(111) surface. The selectivity of glyceraldehyde toward C–H bond scission over C–OH bond scission explains the higher selectivity of 1,2-PDO over 1,3-PDO.