Machine learning predicts and optimizes hydrothermal liquefaction of biomass

The hydrothermal liquefaction process has recently attracted more attention in biorefinery design and imple-mentation because of its capability of handling various wet biomass feedstocks. However, measuring the quantitative and qualitative characteristics of hydrothermal liquefaction (by)products is challenging because of the need for time-consuming and cost-intensive experiments. Machine learning technology can cope with this issue thanks to its ability to learn from past datasets and mechanisms. Hence, machine learning was applied herein to quantitatively and qualitatively characterize hydrothermal liquefaction (by)products based on biomass composition and reaction conditions. The data patterns compiled from the published literature were used to develop a universal machine learning model applicable to a wide range of biomass feedstocks and reaction conditions. The collected data were statistically analyzed and mechanistically discussed. Among the four ma-chine learning models considered, Gaussian process regression could provide the highest accuracy, with a cor -relation coefficient higher than 0.926 and a mean absolute error lower than 0.031. An effort was also made to maximize biocrude oil quantity and quality and minimize byproducts quantity using the objective functions developed by the selected model. The optimal biocrude oil yield (48.7-53.5%) was obtained when the carbon, hydrogen, nitrogen, oxygen, sulfur, and ash contents of biomass were in the range of 40.9-48.3%, 9.72-9.80%, 11.9-13.6%, 15.2-15.6%, 0.0-0.94%, and 0.0-2.92%, respectively. The optimal operating conditions were: operating dry matter = 31.4-33.0%, temperature = 394-400 ?C, reaction time = 5-9 min, and pressure = 30.0-35.6 MPa. An easy-to-use software package was developed based on the selected machine learning model to pave the way for bypassing unnecessary lengthy and costly experiments without requiring extensive machine learning knowledge. The present study highlights the vast potential of machine learning for modeling biomass hydrothermal liquefaction.

Automated summary

This study applies machine learning to quantify and qualify hydrothermal liquefaction products based on biomass composition and reaction conditions. A universal machine learning model is developed using data patterns compiled from published literature, and Gaussian process regression is found to provide the highest accuracy. Optimal operating conditions and objective functions are developed to maximize biocrude oil yield and minimize byproducts. An easy-to-use software package is also developed to bypass costly and lengthy experiments.