Machine learning prediction of physical properties and nitrogen content of porous carbon from agricultural wastes: Effects of activation and doping process.

[Display omitted] • Machine learning were used to predict activation biomass-derived N-doped porous carbon. • Physicochemical properties of porous carbon were well predicted by GBDT model. • The combination with all the features of describing biomass feedstock performed best. • Impregnation ratio and activation temperature were critical to pore structure of porous carbon. • The nitrogen content is related to the active substance exposed by the activator etching. In the study, machine learning models for prediction specific surface area (SSA), total pore volume (TPV), microporosity (MP) and nitrogen content (NC) were developed based on three tree-based models by collecting published experimental data on the agricultural wastes for the preparation of nitrogen-doped porous carbon, in which the physicochemical properties of raw materials and activation conditions were considered and compared as input features. By randomly combining different groups of input features in the three models, gradient boosted decision tree (GBDT) with the combination of all features presented the highest test R2 of 0.939, 0.835, 0.861 and 0.906 for the four prediction targets with high accuracy and good generalization ability. The visual analysis of the model showed that the activation temperature and impregnation ratio were the most critical factors for the targets, while the NC was additionally influenced by the nitrogen-doping ratio. The SSA, TPV and MP peaked around the activation temperature of 800 °C and the impregnation ratio of 3, while NC also increased with the amount of nitrogen doping and reached saturation of nitrogen fixation at the nitrogen-doped ratio of 3. This data-driven scientific tool effectively guided the synthesis of porous carbon and broadened the utilization of biomass energy. [ABSTRACT FROM AUTHOR]