Abstract
Biomass pyrolysis is a versatile thermochemical route that converts lignocellulosic and organic waste into bio-oil, biochar, and non-condensable gas under oxygen-limited conditions, providing a renewable pathway to fuels, chemicals, and carbon-rich solids. Diversifying biomass feedstocks for pyrolysis-based bioenergy requires predictive tools that resolve bio-oil composition at the functional-group level, yet existing models are trained on narrow feedstock ranges and predict only aggregate yields, precluding compound-specific upgrading decisions. No unified framework currently couples process simulation with deployable machine-learning (ML) surrogates for both lignocellulosic and non-lignocellulosic biomasses at the compound-group level. We hypothesize that a simulation-calibrated gradient-boosting surrogate, trained on biochemical-composition features, can generalize bio-oil speciation predictions across chemically heterogeneous feedstocks. We developed an Aspen Plus® V15 RYIELD-based simulation for six feedstocks: bamboo, eucalyptus, pine, corn cob, banana peel, and fish processing waste at 550 °C, generating 594 observations of five bio-oil compound groups and biochar yield. Six supervised ML algorithms were benchmarked on a 70/30 stratified split with cellulose–hemicellulose–lignin mass fractions as continuous predictor features, and the best model was deployed in an interactive Python dashboard. Results show that XGBoost achieved the highest predictive accuracy (R² = 0.951, RMSE = 3.987 kg h⁻¹), outperforming neural network (0.943) and random forest (0.938), while simulation outputs reproduced published experimental compound-group ranges (Pearson r = 0.91, RMSD = 4.8 wt%). Biochar yield spanned 61.87–391.62 kg h⁻¹, correlating with fixed carbon and lignin content. This framework provides the first simulation–ML–dashboard pipeline enabling real-time, compound-resolved feedstock screening for pyrolysis system design. Feature analysis further indicated that biochemical composition, particularly the cellulose–hemicellulose–lignin fractions, governed inter-feedstock yield stratification, with corn cob yielding the highest organic-acid fraction and eucalyptus and fish-processing waste the highest phenolic and carbonyl fractions. The findings of this study can help for better understanding of how feedstock biochemical composition controls compound-group bio-oil speciation, and can support rapid feedstock screening, upgrading-pathway selection, and preliminary process design in multi-feedstock pyrolysis systems without repeated high-fidelity simulation.