Explainable machine learning for unified prediction of activation energy in combustion and pyrolysis of biomass, biochar, and their mixtures

Abstract

This study presents a comprehensive approach, incorporating feature engineering, hyperparameter tuning, and rigorous cross-validation, to develop a robust and accurate model for predicting the activation energy (E a ) in biomass combustion and pyrolysis processes. Four tree-based machine learning models, including decision tree (DT), random forest (RF), gradient boosting decision trees (GBDT), and extreme gradient boosting (XGB), were evaluated using 1,416 data points. RF demonstrated the best performance, with R 2 values of 0.9663 (training) and 0.9081 (testing), and RMSE values of 11.2788 and 18.5770, respectively. SHapley Additive exPlanations (SHAP) analysis identified conversion degree and flow rate as the most influential predictors of E a based on optimized RF model. Furthermore, the RF model consistently outperformed other models on new datasets, achieving low deviations (< 10.12 kJ mol –1 ) from actual E a values and showing reduced susceptibility to overfitting. These findings suggest that RF can serve as efficient tools for predicting E a , offering an alternative to time-consuming experimental procedures in thermogravimetric analysis.