Machine learning models for fast and isothermal hydrothermal liquefaction of biomass: Comprehensive experiment and prediction of various product fraction yields

Abstract

Hydrothermal liquefaction (HTL) shows promise in biocrude productions from wet biomass. In this study, phycocyanin conversion was performed under an extensive range of conditions (set-point temperatures of 300–500 °C and holding times of 20–1800 s), covering both isothermal and fast HTL processes. Based on 226 sets of product fraction yields obtained from this study and literature, machine learning models, such as multiple linear regression, decision regression tree, random forest, gradient boosting regression, and support vector regression were developed. The model with the highest predictive accuracy for biocrude yield was further used to perform both single and multi-task predictions of other product yields. The results show that fast HTL at 500 °C and 40 s produced the highest biocrude yield (35.0 wt%) and energy recovery (55.2%), which were 11.9 wt% and 15.9% higher than those obtained by conventional isothermal HTL at 300 °C and 1800 s. Biocrude predominantly consisted of nitrogenous heterocyclic compounds, leading to its high nitrogen content (∼10 wt%). In addition, water quality analysis shows that the chemical oxygen demand, ammonia and total phosphorus contents of the aqueous phase by-products were much higher than the discharge standards and required further treatments. Cross-validation and testing error analysis show that the random forest model exhibited the highest predictive accuracy for the biocrude yield, with a testing R2 of 0.93. Additionally, the random forest model could predict other product fraction yields with desirable accuracies. According to the feature importance analysis, the most important factor for biocrude production from fast HTL was the residence time, followed by the reaction temperature.