Machine learning-driven prediction and optimization of monoaromatic oil production from catalytic co-pyrolysis of biomass and plastic wastes

Abstract

Listen

Translate article

Catalytic co-pyrolysis of biomass and plastic wastes is an efficient way for monoaromatic-rich oil production, while it is difficult to conclude oil evolution rule due to the complex interaction of multiple factors during co-pyrolysis. Hence, multiple machine learning algorithms were devised to aid process optimization and zeolite catalyst screening. The results showed that gradient boosting decision tree model performed the best prediction accuracy and generalization ability, as its coefficient of determination (R2) and root-mean-square error were 0.90 and 5.04, compared to random forest (R2 ∼ 0.84), extreme gradient boosting (R2 ∼ 0.90), and light gradient boosting machine (R2 ∼ 0.86). Both feature importance and partial dependence analysis showed that operating parameters, catalyst properties, and feedstock composition were in the descending order to affect oil yield with monoaromatic selectivity. The reaction temperature (∼500 °C) and feedstock/catalyst ratio (<5:1) had preferable influence on higher oil yield. The optimal selectivity of benzene, toluene, ethylbenzene and xylene (BTEXs) in oil would be obtained at plastic proportion of around 60 wt% and Si/Al ratio of 20–30. The two-way PDA about interaction effect between multiple factors for oil products were also discussed in detail. The work favors to rationalize large-scale oil production from pyrolysis of solid wastes.