Machine learning aids solvothermal liquefaction of algal biomass: Prediction of nitrogen content and bio-oil yield

Abstract

In order to generate bio-oil with enhanced characteristics, many solvents have been frequently used in the liquefaction of algal biomass. However, the findings differ from one work to another, and no generalization has yet been performed. Various works have reported on the machine learning application to the liquefaction of biomass but none of them offered a specific attention to the impact of the reaction solvent. In this study, the Random Forest (RF) and eXtreme Gradient Boosting (XGB) methods were used to predict the yield (Yb) and nitrogen content (Nb) of bio-oil from solvothermal liquefaction of algal biomass. On unknown test data, XGB provided superior models with R2 and RMSE of (81.61% and 8.08 wt%) and (93.33%, 0.6 wt%) for Yb and Nb, respectively. This study demonstrates the relevance of the solvent's hydrogen-bonding donor strength over its polarity and refractive index in predicting Yb and Nb. The partial dependence plots exhibited the trends of the various inputs’ partial impacts. To facilitate prediction for other researchers, a graphical user interface was developed based on the best models. With this user interface, the researchers could explore the Yb and Nb for various mono-solvents and/or binary solvents and find out the optimized parameters according to the need, before go to experimental works. This will help save time and resources that are usually spent on several trial experiments that may sometimes not yield positive results. This work will instruct researchers for the selection of solvent in order to reach desirable Yb and Nb.