Machine learning technologies for digestate-derived hydrochar yields

Digestate is a promising precursor for hydrochar production, while composition complexity makes hydrochar engineering challenging. Machine learning (ML) is a cost-effective tool for modeling hydrochar, although the ML model specialized in digestate biomass is lacking. For the first time, this work adopted ML techniques to predict the digestate-derived hydrochar yield, with the dataset only including digestate as biomass resources. Random forest (RF) and eXtreme Gradient Boosting (XGB) predicted the yields based on digestate elemental/proximate compositions and HTC reaction parameters. RF and XGB performed satisfactorily, with test R² achieving 0.907 and 0.911, respectively. HTC parameters are the dominant factors affecting yield prediction. Inspired by the ML results, the new modified severity factor (Ro'), integrating the effect of temperature, time, and solid loading, was proposed for modeling hydrochar yield. The proposed Ro' factor showed better predictive performance and generalizability than the conventional severity factor (Ro), with improved R² (0.822–0.978–0.885–0.984) and reduced average absolute prediction bias (2.56–1.71 %). The proposed Ro' model further addresses data deficiency issues, significantly improving the CV scores (mean R²) of RF (0.703–0.851) and XGB (0.791–0.905), suggesting data augment mitigates the overfitting tendency. HTC parameters, essential for scale-up, including reactor size, stirring rate, particle size, and the catalyst effect, should be included as training features in future work to enhance model predictive performance.

© 2025 Institution of Chemical Engineers. Published by Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.