Machine-learning-aided biochar production from aquatic biomass

Modeling hydrothermal carbonization (HTC) and pyrolysis carbonization (PLC) for the conversion of biomass into high-quality biochar for various applications shows promise. Unlike the extensive modeling studies on lignocellulosic biomass, research on aquatic biomass (AB) had not been reported until now. In this study, we compiled 586 data points from existing literature and trained five tree-based models to predict the yields of hydrochar and pyrochar and their properties, including nitrogen recovery degree, energy density, energy recovery degree, and residual sulfur degree, based on 10 feedstock and process parameters. The random forest regression (RFR) model demonstrated the highest predictive accuracy among these models. It achieved R² values ranging from 0.89 to 0.98 for hydrochar yield, nitrogen recovery degree of hydrochar, energy recovery degree of hydrochar, and residual sulfur degree of hydrochar. The extreme gradient boosting (XGB) model also showed exemplary performance, with R² values between 0.84 and 0.94 for energy density of hydrochar, pyrochar yield, and nitrogen recovery degree of pyrochar. Results on feature importance highlighted that, beyond the well-documented impact of process parameters, the properties of biochar were significantly influenced by the elemental compositions, such as nitrogen and sulfur contents of the feedstock. The relationship between these factors was further elucidated using partial dependence plots. Finally, we used RFR model for hydrochar yield and XGB model for pyrochar yield as examples, to test generalization ability of developed models with new data, further explaining their application methods. Overall, this study provided valuable insights into predicting and understanding the HTC and PLC processes of AB to produce high-quality biochar for various applications using low resources and time costs. Besides, we presented an iterative learning application method where the developed models demonstrated exceptionally high performance with new data. This method is highly versatile and can be adopted across various directions in the field of machine learning. Graphical Abstract: (Figure presented.)