Interpretable modeling of biomass fractionation under acidic pretreatment via multi-step data augmentation and an...

Acidic pretreatment is a key step in biomass refining for the targeted separation of major lignocellulosic components. However, nonlinear interactions among multiple process variables and acid structural characteristics complicate the interpretation of separation behavior and the prediction of pretreatment outcomes. In this study, a machine-learning database containing 142 literature-derived records was constructed by integrating acid structural descriptors, physicochemical properties, and process parameters. To address the challenges of small sample size, uneven data distribution, and sparse records under extreme operating conditions, a training-set-only data augmentation strategy was adopted. Minority-sample synthesis, local perturbation, and residual-based augmentation were used to improve sample coverage in sparse regions, while audits based on sample IDs, lineage IDs, and parent-sample sources were performed to ensure that the test set remained original and independent. An entropy-weighted Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS) method was further introduced to construct target-specific weighted ensemble models. Strict test-set evaluation showed that the ensemble models achieved R² values of 0.476, 0.564, and 0.804 for hemicellulose, cellulose, and lignin separation rates, respectively. These results indicate that lignin separation exhibited more stable predictability, whereas the prediction of hemicellulose and cellulose separation was more strongly constrained by small-sample heterogeneity. Partial dependence plots and accumulated local effect analyses showed that temperature, acid concentration, and reaction time were the core variables controlling component separation rates, while acid structural and functional-group features mainly acted as secondary modulators. Overall, this study establishes a data-driven framework that integrates leakage prevention, robust prediction, and interpretable analysis, providing a useful approach for screening acidic pretreatment windows and generating mechanistic hypotheses regarding component separation behavior.