From machine learning benchmarking to process design: Integrated prediction of biochar yield and specific surface area with hyperparameter optimization and GUI deployment

Biochar is a sustainable carbon-rich material with applications in carbon sequestration, soil improvement, environmental remediation, and renewable energy. Accurate prediction of biochar yield and specific surface area is essential for optimizing feedstock selection and pyrolysis conditions while reducing costly experimental trials. This study developed and benchmarked six supervised machine learning models—XGBoost, LightGBM, CatBoost, Random Forest, AdaBoost, and Artificial Neural Network (ANN)—for integrated prediction of biochar yield (%) and specific surface area (m²/g) using ten feedstock composition and pyrolysis process variables. Datasets comprising 558 observations for biochar yield and 344 for specific surface area were analyzed. XGBoost achieved the highest predictive performance, with test R² of 0.948 (RMSE = 3.46%) for yield and 0.909 (RMSE = 33.34 m²/g) for specific surface area. LightGBM, CatBoost, and Random Forest showed competitive performance, whereas ANN and AdaBoost were less accurate. The lower accuracy for specific surface area reflects the greater complexity of pore structure development compared with bulk biochar formation. Bayesian optimization and randomized search combined with five-fold cross-validation improved model robustness and generalization. SHAP analysis identified heating rate, temperature, ash content, and engineered thermal features as dominant factors influencing biochar yield, while specific surface area (SSA) was primarily governed by pyrolysis conditions and feedstock-related properties. Interaction analysis further revealed nonlinear relationships among key features, consistent with the complex mechanisms governing biochar yield and porosity development during pyrolysis. A user-friendly graphical user interface (GUI) enabled rapid single-sample and batch prediction without programming expertise. The proposed framework provides an accurate, interpretable, and deployable decision-support tool for feedstock screening and pyrolysis optimization, supporting sustainable biomass-feedstock valorization and carbon sequestration.