Machine learning prediction of hydrochar properties using ash-based biomass classification with optimized models and interpretability analysis

The growing demand for sustainable energy solutions has intensified research on biomass valorization into carbon rich hydrochar via hydrothermal carbonization (HTC). However, optimizing hydrochar fuel properties (Fuel Ratio, H/C, O/C) remains challenging due to the complex interplay of feedstock composition and process conditions. This study presents a machine learning (ML) driven approach to predict hydrochar properties by categorizing biomass into low, medium and high ash content groups. An integrated framework combining different ML models and metaheuristic optimization techniques was developed to establish robust predictive capabilities. The results show that the GPR model achieved a testing R 2 > 0.72 and RMSE < 0.1365 across all ash categories. Feature selection highlighted temperature, time, and biomass composition (Carbon, Oxygen, Volatile Matter, Fixed Carbon) as important predictors, demonstrating the effectiveness of ML with hyperparameter optimization for accurate hydrochar properties prediction. Partial dependence plots further explained nonlinear relationships between inputs and outputs. A user-friendly Graphical User Interface was developed, enabling real-time hydrochar properties prediction, closely aligning with experimental data. This work demonstrates that ML models, coupled with optimization techniques, can effectively guide HTC process optimization for tailored hydrochar production, bridging the gap between data-driven insights and practical applications in renewable energy.