Role of glycosidic bond in initial cellulose pyrolysis: Investigation by machine learning simulation

Abstract

Pyrolysis of cellulose has been widely investigated based on experiments and density functional theory (DFT) calculations, whereas the role of glycosidic bonds remains unclear. In this work, the machine learning based method has been employed in cellulose pyrolysis mechanism study, with β-d-glucose and methyl β-d-glucoside as model compounds. The initial pyrolysis reactions are automatically determined through potential energy surface (PES) searching based on the combined methodology of stochastic surface walking and neural network potential (SSW-NN). 229 and 43 reaction pairs with reliable transition states have been obtained in the reaction space of β-d-glucose and methyl β-d-glucoside, respectively, and 33 elementary reactions in the initial pyrolysis stage have been found among these reaction pairs. Besides the well-known reactions, new and feasible reactions are identified by the SSW-NN method. Particularly, it is newly found that the ring-opening and glycosidic bond breaking involve multiple hydrogen transfer processes, where the hydrogen transfer is assisted by 3-OH and/or 6-OH. Hence, these reactions have relatively low energy barriers. The reactions involving the participation of the C1 group are significantly influenced by the glycosidic bond. On the contrary, other reactions without involving the C1 group are almost not affected by the glycosidic bond. Such results should be ascribed to the distinction of electrostatic potentials of the C1 free hydroxyl and the glycosidic bond. This study can provide new insight into the cellulose pyrolysis mechanism based on the SSW-NN method.

Keywords

• Cellulose
• Glycosidic bond
• Machine learning
• SSW-NN
• DFT