Machine learning prediction of delignification and lignin structure regulation of deep eutectic solvents pretreatment processes

Abstract

Prediction of the pretreatment efficiency of lignocellulosic biomass with ternary deep eutectic solvents (DES) containing Lewis acids by machine learning (ML). Principal component analysis, partial least square method, spearman correlation matrix, random forest, extreme gradient boosting and deep neural network were used to elucidate the correlation between 77 variables and the mechanism of lignin depolymerization. The effects of raw material composition, reaction conditions, physicochemical properties of DES and structural parameters in lignin on 9 target variables including β-O-4 bond, β-β bond, β-5 bond, weight average molecular weight, number average molecular weight, polydispersity index, ratio of syringyl units to guaiacyl units, content of phenolic hydroxyl groups and delignification were analyzed. Multivariate analysis showed that temperature, polarity related parameters of HBD and acidity of Lewis acids contributed significantly to the degree of lignin depolymerization. The types and fracture mechanisms of the bonds between different structural units of lignin can be determined by the analysis of structural parameters. XGBoost model has the best performance among all the ML models, and the R square of the test sets for the target variables is above 0.76. Feature importance analysis showed that structural parameters significantly affected the pretreatment effect. The physical and chemical parameters of HBD, such as dipole moment, Log P and surface tension should be paid attention to in the design of DES. The study of the weak intermolecular forces in the lignin and DES systems is beneficial to reveal the mechanism of the pretreatment process. This study provides novel insights into the structural regulation and high-value utilization of lignin in the process of DES pretreatment of lignocellulosic biomass.