The Key Factors of Separation of Biomass Components in Acid Pretreatment Environment Were Studied by Machine Learning

Abstract

In this study, machine learning (ML) tools were used to investigate the impact of key variables in acid pretreatment on the separation efficiency of the three major macromolecules in lignocellulosic biomass. Methods such as Principal Component Analysis (PCA), Partial Least Squares (PLS), Spearman's correlation matrix, Random Forest (RF), and Backpropagation Neural Networks (BPNN) were applied to analyze the correlations between 17 variables and the separation patterns of cellulose, hemicellulose, and lignin. Machine learning models, including RF, XGBOOST, and BPNN, established key relationships between feedstock composition, reaction conditions, functional group types, solvent molecular structures, and separation rates. The results showed that the RF model was most suitable for the acid pretreatment system, with reaction temperature and the hydroxyl content in the solvent significantly influencing lignin separation. By combining PCA, PLS, Spearman's analysis, and the RF model, it was found that increasing the number of hydroxyl groups in the solvent and maintaining an appropriate reaction temperature were crucial for lignin separation. This research provides valuable insights into the structural regulation of macromolecule separation during the acid pretreatment of lignocellulosic biomass, contributing to the scalable application of acidic pretreatment technologies.

Keywords: Biomass, Machine Learning, Acid Solvent, Pretreatment condition, Functional group