Machine learning-driven predictions of biochar yield and NPK composition: insights into biomass pyrolysis with data augmentation and model interpretability

Abstract Biochar's potential as a sustainable solution for agricultural and environmental management depends on its capacity to retain nutrients and sequester carbon. However, accurately predicting biochar yield and nutrient content, particularly nitrogen (N), phosphorus (P), and potassium (K), remains a significant challenge. This study addressed this issue by applying advanced machine learning models to predict biochar properties based on biomass characteristics and pyrolysis conditions. The models included Support Vector Regression (SVR), Random Forest (RF), Back Propagation Artificial Neural Network (BP-ANN), and Extreme Gradient Boosting (XGBoost). Analysis of 271 datasets, augmented with random noise injection for data augmentation, revealed that XGBoost was the most reliable model, achieving an average R 2 of 0.97 for predicting biochar yield and elemental compositions. Key findings indicate that pyrolysis temperature is the primary determinant of biochar yield, while feedstock composition plays a critical role in nutrient retention. Additionally, a novel graphical user interface (GUI) was developed to translate these computational insights into practical applications, bridging the gap between complex data analysis and real-world agricultural and environmental management. This research offers a robust, data-driven framework for optimizing biochar production and enhancing its role in sustainable agriculture and environmental conservation. Graphical Abstract