Constru of lignin model molecule based on deep learning and infrared spectroscopy

Abstract

A deep learning-based approach, LigninGen, is presented for the construction of model molecules for lignin. It is able to accurately predict the molecular structure of experimental lignin samples through multi-stage pretraining and zero-shot prediction. The model combines the results of IR, ¹³C NMR, and ultimate analysis, and the structural characteristics of the model molecules are highly consistent with the experimental data. Moreover, LigninGen utilizes SELFIES encoding to directly generate the 3D coordinates of the molecules, which is a more realistic reflection of the amorphous structure of lignin compared to the traditional methods.