Integration of Random Forest Feature Importance and Machine Learning Frameworks for Prediction of Specific Surface Area of Biochar Derived from Rice Residues
Аннотация
The specific surface area (SSA) of biochar is a key property influencing its performance in soil amendment, wastewater treatment, and carbon sequestration. This study aimed to develop a hybrid data-driven framework integrating Random Forest-based feature importance with advanced machine learning (ML) models to predict the SSA of biochar derived from rice residue pyrolysis. Historical pyrolysis data were compiled from the literature, and missing values were treated using multiple imputation techniques, whose impact on model performance was systematically evaluated. Random Forest-based feature importance was applied to identify the dominant predictors as temperature, fixed carbon, and volatile matter. These predictors were used to develop Multi-Expression Programming (MEP), Gradient Boosting Regression Trees (GBRT), and Gaussian Process Regression (GPR) models. The MEP, GBRT, and GPR models have R 2 values of 0.910, 0.932, and 0.909, respectively, for the testing set. The GBRT model achieved the highest predictive performance, while MEP provided explicit mathematical expressions describing the nonlinear relationships between process variables and SSA. Cross-model consistency between the MEP sensitivity analysis and GBRT feature importance showed similar model-based rankings of the selected predictors. Although the models were developed using heterogeneous literature data from rice residues, the proposed framework demonstrates an effective and interpretable approach for predicting biochar SSA and can support the optimization of biomass pyrolysis processes.