Multi-Scale Modeling of Enzymatic Hydrolysis Efficiency Based on Machine Learning: From Hydrothermal Pretreatment Conditions to Lignin Structural Properties

Xie D.; Ma H.; Wang J.; Cárdenas-Oscanoa A.J.; Wang S.; Huang C. · ACS Sustainable Chemistry and Engineering (2026) · DOI: 10.1021/acssuschemeng.6c01064

Resumen

The increasing depletion of fossil resources and escalating environmental challenges highlight the urgent need for efficient utilization of lignocellulosic biomass. However, the inherent recalcitrance of lignocellulose hampers its enzymatic degradation and conversion efficiency, making pretreatment essential to improve enzyme accessibility and promote carbohydrate release. In this study, we developed predictive models of enzymatic hydrolysis based on hydrothermal pretreatment by integrating multidimensional features, including feedstock composition, process parameters, and pretreatment outcomes. Among multiple machine learning algorithms, XGBoost demonstrated superior performance across training, testing, and validation sets (R2 = 0.994, 0.841, and 0.856, respectively). At the microscopic level, lignin was further examined as a key factor. The correlation models were established between lignin physicochemical properties, including molecular weight (Mw, Mn), interunit linkages (β-O-4, β-5), functional groups (phenolic, aliphatic, and carboxyl hydroxyl groups), and enzymatic conversion efficiency. XGBoost again demonstrated as the best model, which interpretability was enhanced using feature importance ranking, SHAP dependency analysis, and two-dimensional partial dependence plots (2D-PDP), revealing cross-scale, multidimensional mechanisms within the pretreatment system. By integrating macro-level regulation and microlevel mechanistic models, this study provided a systematic framework for understanding biomass hydrolysis and offers data-driven guidance for optimizing pretreatment strategies and improving conversion efficiency.

Instituciones: Universidad Nacional Agraria La Molina