Machine learning-based prediction of hydrochar higher heating value from process conditions and ultimate & proximate analyses: An alternative to bomb calorimetry?

• ML-based models were investigated as an alternative to bomb calorimetry. • HHV data determined via bomb calorimetry were mined in the literature. • Properties of different hydrochar-derived feedstocks were collected from the literature. • Four ML models (PLSR, RF, SVR, AdaBoost) investigated to predict HHV • RF outperformed during training, while SVR performed better during validation. • SHAP-based feature analysis provided local explanation and interaction effect. Higher Heating Values (HHV) is a very important parameter for analysing fuel properties of resulting materials obtained from Hydrothermal Carbonization (HTC): hydrochar. HHV is determined experimentally by a bomb calorimeter or using empirical correlations only valid for limited hydrochars derived from specific feedstocks and fail to be generalized. Machine Learning (ML) methods are emerging and can give accurate estimates of HHV for multiple feedstock sources. In this study, Partial Least-Squares Regression (PLSR), Random Forest (RF), Support Vector Regression (SVR) and Adaptive Boosting (AdaBoost) are employed to predict hydrochar’s HHV from elemental, proximate data and HTC process conditions. RF outperformed during training with an R 2 and RMSEC of 0.97 and 0.72 MJ/kg respectively, whereas during validation, SVR had the best performances with an R 2 and RMSEP of 0.94 and 1.07 MJ/kg. From Shapley Additive Explanations (SHAP)-based feature analysis, carbon (C), fixed carbon (FC) and ash were the most significant feature contributors for PLSR and AdaBoost, whereas oxygen (O) along with C and FC were found to be highest contributors for the RF model. Finally, regarding SVR, while C is still among the most important features, ash and volatile matter (VM) are also underpinned. By systematically comparing ML models and applying feature analysis, this study not only achieves high predictive accuracy but also provides mechanistic insights into the physicochemical drivers of hydrochar fuel quality, offering a cost-effective alternative to calorimetric experiments. The practical implications of this work include facilitating HTC process optimization, particularly in emerging countries where access to calorimetric equipment is limited.