Interpretable machine learning model for activation energy prediction based on biomass properties

A thorough understanding of pyrolysis kinetics for biomass is crucial for the process design, feasibility assessment, and scale-up in industrial scenarios. To reduce the time and economic cost invested in obtaining experimental kinetic data by performing the thermogravimetric analysis, a data-driven machine learning method was proposed in the present work. Random forest (RF), gradient boosting decision tree (GBDT), and extreme gradient boosting (XGB) algorithms were utilized to predict the activation energy versus conversion range with ten input features from biomass characteristics. Among these XGB has shown the most competitive prediction accuracy with a coefficient of determination (R2) higher than 0.98 and a root mean squared error (RMSE) = 7.7 kJ/mol. To enhance the model interpretability of the derived model, a visualized analysis for individual impact and interactions of variables was conducted by Shapley additive explanation (SHAP). The top three significant inputs for the model established by the current dataset are conversion rate, volatile content, and moisture content. This work provides a quick and accurate kinetic prediction method for unexploited biomass and favors detailed insights into the complicated associations between biomass characteristics and pyrolysis kinetics.