Data-driven machine learning with SHAP for predicting multiple biochar properties from agricultural and forestry waste

Abstract

The pyrolysis of agricultural and forestry waste for biochar production is an effective method for resource utilization and carbon sequestration. The diversity of these wastes leads to variations in biochar properties, limiting its applications. In recent years, machine learning (ML) models have been developed to predict biochar properties. However, there is currently a lack of a generalized predictive model capable of simultaneously evaluating multiple properties of biochar. This study systematically evaluated the performance of six ML models—CatBoost, LightGBM, NGBoost, XGBoost, RandomForest, and AdaBoost—using 288 pyrolysis datasets from 43 types of agricultural and forestry waste to predict biochar yield (BY), carbon content (C), fixed carbon content (FC), ash content, energy yield (EY), higher heating value (HHV), pH, and cation exchange capacity (CEC). Among these models, CatBoost demonstrated the best performance, with R² values ranging from 0.76 to 0.95. Shapley Additive Explanations (SHAP) analysis identified pyrolysis temperature as a key factor influencing biochar yield, carbon content, and pH, while feedstock composition (e.g., cellulose and lignin) was found to significantly affect CEC and FC. The CatBoost model could be used for the dynamic adjustment of pyrolysis parameters based on feedstock, therefore, a modular visual prediction tool was developed using CatBoost for the large-scale production of specific biochar for specific applications by allowing the precise selection of corresponding types of feedstock and pyrolysis conditions.