Hydrochar production from agricultural wastes from Türkiye and predictive modeling: Using RSM and machine learning

Hydrothermal carbonization (HTC) has been widely applied for biomass conversion, comparative studies on diverse agricultural wastes, particularly those originating from Türkiye, remain limited. İntegrated modeling approaches combining response surface methodology (RSM) and machine learning (ML) for predicting hydrochar fuel properties are still insufficiently explored. This study aims to systematically produce hydrochars from walnut shell, hazelnut shell, apricot kernel shell, and olive pomace via HTC and to evaluate the effects of temperature (175–275°C) and residence time (30–60 min) on hydrochar yield and higher heating value (HHV). In addition, predictive models based on RSM and ML approaches were developed to estimate hydrochar properties. Hydrochar yields ranged from 32% to 81%, and HHV increased with increasing process severity, reaching values comparable to low-rank coals. H/C and O/C ratios decreased significantly with temperature and time, confirming enhanced carbonization. RSM models described the response of HHV to temperature and holding time with high accuracy (R2 ≥ 0.94), while MLP models estimated HHV from proximate and ultimate analysis data. A desirability-function optimization of hydrochar yield, energy densification, energy yield, fuel ratio, and carbon retention showed that, except for apricot kernel shell, the conditions that maximize HHV are not those that maximize overall hydrochar quality. Olive pomace-derived hydrochars showed distinct behavior due to higher extractives (lipophilic compounds), resulting in lower carbonization efficiency at mild conditions. Overall, this study fills critical gaps in comparative HTC performance of Turkish agricultural residues and demonstrates the effectiveness of integrated experimental–data-driven approaches for hydrochar property prediction.