Machine-learning-aided hydrochar production through hydrothermal carbonization of biomass by engineering operating parameters and/or biomass mixture recipes

Hydrochar serves not only as a fuel source but also as a versatile carbon material that has found extensive application across various domains. The application performance of hydrochar, e.g., energy recovery and carbon stability, is substantially influenced by its mass yield, higher heating value (HHV), and compositions (C, H, O, N, S, and ash), so the prediction and engineering of these properties is promising. In this study, two machine learning algorithms, namely gradient boosting regression (GBR) and random forest (RF), were used to predict the hydrochar properties mentioned above. The GBR models (with test regression coefficient (R2) values of 0.87–0.98 for single-target prediction and average test R2 of 0.93 for multi-target prediction) exhibited superior predictive capabilities to the RF models (with test R2 of 0.78–0.97 for single-target and average test R2 of 0.90 for multi-target prediction). The interpretation of ML models revealed the importance ranking of features for all targets. Then, engineering hydrochar was carried out through three different optimizations to the as-built multi-target prediction model: i) optimizations of HTC conditions for given biomass samples; ii) optimization of biomass mixture recipes; iii) simultaneous optimization of both biomass mixing recipes and HTC conditions.