Machine learning and SHapley Additive exPlanations to predict product characteristics from coal and plastic co-pyrolysis

Coal co-pyrolyzed with plastic provides a valuable method for producing high-quality liquid fuels, offering the potential reductions in fossil fuel dependence. However, this thermochemical conversion process involves the complex interactions between feedstock properties and operational parameters, making the exploration of this process require a large number of experiments. This study predicts the three-phase yields of coal-plastic co-pyrolysis with four advanced machine learning (ML) algorithms: Artificial Neural Network (ANN), Gradient Boosting Decision Tree (GBDT), Random Forest (RF), and Support Vector Machine (SVM). The feature importance analysis of the optimal model was interpreted through SHapley Additive exPlanations (SHAP) to clarify the input–output relationships. The main results indicate a significant positive correlation between plastic blending ratio and the tar yields, and the chlorine content in plastic was positively correlated with pyrolysis gas yield. Among the four ML models, the ANN achieved superior predictive performance, with the training coefficient of determination (R²) values exceeding 0.94 for all output phases and the testing R² ranging from 0.93 to 0.99. The error metrics confirmed the robustness of ANN, with the training root mean square error (RMSE) and mean absolute error (MAE) of 1.02–3.35 and 0.41–1.19, respectively, while the testing phase errors remained within 2.11–3.47 (RMSE) and 1.56–2.65 (MAE). In addition, the SHAP analysis shows that the plastic blending ratio, pyrolysis temperature and feedstock characteristics are the primary factors during the co-pyrolysis process of coal-plastic. This research provides valuable insights into the co-pyrolysis of coal and plastic and paves the way for further in-depth study.