Interpretable machine learning for prediction and regulation of bio-oil production, quality, and nitrogen content

Bio-oil derived from biomass pyrolysis is a promising renewable liquid fuel, however, its utilization is limited by unstable production, low heating value, high acidity, and the presence of nitrogen-containing compounds. In this study, an interpretable machine learning framework was developed to predict and analyze four bio-oil production- and quality-related properties: yield, higher heating value (HHV), pH, and nitrogen content. A literature-derived dataset comprising 430 experimental records from 81 biomass types was compiled, covering feedstock characteristics and pyrolysis conditions. Three tree-based models, including Light Gradient Boosting Machine, Random Forest, and Extreme Gradient Boosting, were compared. LGBM showed the best overall performance, with R 2 values of 0.9228, 0.9673, 0.9059, and 0.9328 for bio-oil yield, HHV, pH, and nitrogen content, respectively. SHAP analysis identified oxygen, carbon, fixed carbon, ash, and nitrogen as key factors governing the evolution of bio-oil properties. Oxygen and carbon primarily controlled HHV; oxygen and nitrogen regulated acidity; fixed carbon and ash affected liquid-product formation; and feedstock nitrogen was the dominant factor affecting bio-oil N. Temperature-dependent interaction analysis further indicated that moderate pyrolysis temperatures were associated with higher bio-oil yield and bio-oil N levels, whereas excessive temperatures enhanced secondary cracking, deoxygenation, and possible nitrogen redistribution. These findings provide data-driven guidance for biomass feedstock selection, optimization of pyrolysis conditions, improvement of bio-oil production, and quality regulation.