Hybrid statistical and machine learning optimization of continuous pyrolysis to maximize bio-oil production

Process optimization is a crucial step in biomass conversion to maximize product yield and control product properties, particularly for bio-oil production. In this study, a hybrid statistical and machine learning (HSML) approach, combining response surface methodology (RSM) with Gaussian process regression (GPR) and support vector regression (SVR), was applied to maximize bio-oil yield during continuous pyrolysis using local common reed as feedstock. Quadratic model fitting and statistical analysis via RSM were first discussed, followed by the hyperparameter tuning and interpretation of machine learning (ML) algorithms. Comparison of predicted and experimental yields confirmed the reliability and feasibility of the HSML approach. The optimal conditions were identified as temperature at 490 °C, carrier gas flow rate at 220 SCCM, and feeding rate at 10.8 RPM (1.5 g min−1), with predicted bio-oil yields of 39.82 wt% (RSM), 39.84 wt% (GPR), and 39.97 wt% (SVR), respectively. The validation experiments under optimal conditions also yielded a comparable result of 39.90 ± 0.35 wt%. The bio-oils, including light and heavy fractions, obtained from optimal conditions were characterized, including elemental composition, functional group, and chemical composition. Light bio-oil showed high oxygen content (72.36 wt%) and was rich in oxygenated compounds, especially acids (31.01 %) and aldehydes (16.97 %). Conversely, heavy bio-oil contained more carbon (61.70 wt%) with phenolic compounds (41.37 %) dominating. Besides, particle size was found to influence both yield and oil quality, where smaller particles (<250 μm) increased bio-oil yield (40.40 wt%) and reduced large oxygenates in the light fraction. Overall, the findings demonstrate HSML as a powerful tool for pyrolysis optimization and bio-oil maximization.