Machine learning-aided prediction of bio-BTX and olefins production from zeolite-catalyzed biomass pyrolysis

In this work, the yields of bio-BTX (biomass-derived benzene, toluene and xylene) and light olefins from zeolite-catalyzed biomass pyrolysis were predicted for the first time via machine learning methods based on various experimental parameters, including biomass characteristics, catalyst properties and pyrolysis conditions. Random forest (RF), gradient boosting decision tree and extreme gradient boosting algorithms were used for evaluation, and the results reveal that the RF algorithm has the highest prediction accuracy with the high test R2 values (0.91–0.97) for all models. The importance and correlation between experimental parameters and output targets were clearly identified by feature importance analysis, for example, the pyrolysis conditions have significant effect on the yields of benzene and light olefins, and the yields of toluene and xylene were mainly affected by catalyst properties (surface area and Si/Al ratio) and biomass characteristics (cellulose and hemicellulose contents), respectively. Furthermore, the partial dependence analysis (PDA) was evidenced to be powerful to understand the influence of input features on the output targets and the interaction between input features. The results will provide a feasible reference for the selection of parameters and prediction of target product yields in zeolite-catalyzed biomass pyrolysis.