Predicting the chemical composition of biocrude from hydrothermal liquefaction of biomasses using a multivariate statistical approach

Hydrothermal liquefaction (HTL) is a promising technique for the conversion of wet biomasses into a complex biocrude. In this study, biocrudes from 10 different feedstocks of Spirulina, Miscanthus, sewage sludge and their mixtures were investigated in a mixture design. Furthermore, the effects of temperature (250-350 degrees C), reaction time (5-31 min), and solid loading (5-25 wt%) were investigated using a central composite design. Analysis of the biocrudes was performed using gas chromatography coupled to mass spectrometry (GC-MS). The software PARADISe was applied to deconvolute chromatographic peaks and tentatively identify 152 compounds, including small carboxylic acids, fatty acids, hydrocarbons, alcohols, carbonyls, amino acids, carbohydrates, oxygenated aromatics and nitrogen-containing compounds. Principal component analysis (PCA) separated the samples corresponding to feedstock in PC1 and PC2, whereas PC3 separated samples based on their process conditions. Partial Least Squares (PLS-R), random forest, lasso, ridge, and gradient boosting regressors were applied to develop predictive models and their performance was compared. The models were evaluated according to the coefficient of determination (R2), root mean square error (RMSE), and bias values. This work highlights the differences in biocrudes from HTL of feedstocks of varying biochemical composition and presents new knowledge of the effect of biochemical composition and process conditions on different compound classes found in the biocrude. The results thus provide valuable information for the optimization of biocrude production via HTL. The effects of hydrothermal liquefaction process conditions and biochemical composition of feedstock on biocrude yield and major compound classes were investigated using regression models and machine learning.