Prediction of phenol yield by machine learning based on biomass characteristics, pyrolysis conditions, and catalyst properties

Phenol is one of the most valuable chemicals from biomass pyrolysis. The study of phenol production is time-consuming and uneconomical by traditional experiments. In this study, random forest (RF) and extreme gradient boosting (XGB) were used to predict phenol yield based on biomass characteristics, pyrolysis conditions, and catalyst properties. The results indicated the XGB model had a better prediction performance (optimal test R² = 0.93). The Shapley additional exploration showed that the lignin content (Lig) and the ratio of catalyst to biomass (C/B) had more impact on phenol yield. The one-dimensional partial dependency plots suggested that phenol yield first increased with the increase of Lig (Lig < 35 wt%) and C/B (C/B < 1), and then decreased with the rise of Lig (Lig > 35 wt%) and C/B (C/B > 1). The two-dimensional partial dependency plots indicated that the highest phenol yield could reach 66 mg/g (Lig ≈ 43 wt% and PT ≈ 570 °C). Moreover, the predictive performance of the model was verified by experiments. All prediction errors were within ± 10 %, achieving higher accuracy. This study provides a convenient and economical way to evaluate and optimize pyrolysis experiments to improve phenol yield and provides a scientific reference for efficient utilization of biomass and bio-oil production.