A dual-stage machine learning framework for comprehensive characterization of waste plastic pyrolysis oils

Abstract

Pyrolysis is a promising approach to recycling plastic waste by converting it into valuable fuels and chemicals. However, process optimization is hindered by complex interdependencies among critical factors (e.g., plastic composition, pyrolysis temperature, and catalyst acidity) and empirical limitations (e.g., small datasets, nonlinear dependencies, and insufficient mechanistic understanding of composition-specific pathways) in predicting product distributions. To address these challenges, this study proposes a dual-stage machine learning framework for the comprehensive characterization of waste plastic pyrolysis oils, including their yield prediction and detailed compositional analysis. In the first stage, four single models (DT, RF, SVR, and XGB) were employed to predict oil yields, with XGB achieving superior accuracy (R² = 0.959) and SHAP analysis identifying pyrolysis temperature and catalyst acidity as dominant influencing factors. The second stage introduced a stacking-based fusion model to predict detailed oil compositions (alkanes, alkenes, and aromatics), outperforming single models by effectively capturing high-dimensional nonlinearities and feature interactions. The catalyst properties are the most influential for composition, with partial dependence plots delineating optimal conditions (e.g., 530°C for alkanes and 620°C for aromatics) and suggesting acid-catalyzed conversion of alkenes to aromatics. By bridging empirical correlations with mechanistic insights, this framework enables targeted pyrolysis optimization for high-value oil production. The study advances plastic waste valorization through integrated yield and compositional analysis, providing a transformative tool for sustainable resource recovery and circular economy applications.