Prediction and optimization design of porous structure properties of biomass-derived biochar using machine learning methods

Biochar produced from biomass by pyrolysis and activation is a platform porous carbon material that has been widely used in many areas. The porosity properties of biochar such as specific surface area (SSA), total pore volume (Total_PV), micropore volume (Micro_PV), mesopore volume (Meso_PV), and average pore size (Average_PS) are essential to biochar applications. Although previous machine learning (ML) models can precisely predict SSA and Total_PV, these models are unable to comprehensively predict the other porosity characteristics. More importantly, activation, which is a critical process for preparing high-porosity biochar, was generally not considered in previous studies. Here, six single-target models were established first based on pyrolysis & activation conditions for the prediction of SSA, Total_PV, Micro_PV, Meso_PV, Average_PS, and yield, obtaining test R² of 0.89, 0.86, 0.88, 0.89, 0.76 and 0.91, respectively. Then, a multi-target model was established for simultaneous prediction with an average test R² of 0.87. ML model interpretation indicated agent type and ratio were crucial to porosity properties. Finally, activation and direct pyrolysis biochar production optimum schemes were derived from ML model for a high porosity. Favorable experimental verification results were obtained with validation R² of 0.98, indicating the great potential of using ML for biochar engineering.