Precise Prediction of Biochar Yield and Proximate Analysis by Modern Machine Learning and SHapley Additive exPlanations

Biochar is found to possess a large number of applications in energy and environmental areas. However, biochar could be produced from a variety of sources, showing that biochar yield and proximate analysis outcomes could change over a wide range. Thus, developing a high-accuracy machine learning-based tool is very necessary to predict biochar characteristics. In this study, a hybrid technique was developed by blending modern machine learning (ML) algorithms with cooperative game theory-based Shapley Additive exPlanations (SHAP). SHAP analysis was employed to help improve interpretability while offering insights into the decision-making process. In the ML models, linear regression was employed as the baseline regression method, and more advanced methodologies like AdaBoost and boosted regression tree (BRT) were employed. The developed prediction models were evaluated on a battery of statistical metrics, and all ML models were observed as robust enough. Among all three models, the BRT-based model delivered the best prediction performance with R2 in the range of 0.982 to 0.999 during the model training phase and 0.968 to 0.988 during the model test. The value of the mean squared error was also quite low (0.89 to 9.168) for BRT-based models. SHAP analysis quantified the value of each input element to the expected results and provided a more in-depth understanding of the underlying dynamics. The SHAP analysis helped to reveal that temperature was the main factor affecting the response predictions. The hybrid technique proposed here provides substantial insights into the biochar manufacturing process, allowing for improved control of biochar properties and increasing the use of this sustainable and flexible material in numerous applications.