Machine learning prediction of nitrogen heterocycles in bio-oil produced from hydrothermal liquefaction of biomass

Abstract

Hydrothermal liquefaction (HTL) of high-moisture biomass or biowaste to produce bio-oil is a promising technology. However, nitrogen-heterocycles (NH) presence in bio-oil is a bottleneck to the upgrading and utilization of bio-oil. The present study applied the machine learning (ML) method (random forest) to predict and help control the bio-oil NH, bio-oil yield, and N content of bio-oil (N_oil). The results indicated that the predictive performance of the yield and N_oil were better than previous studies, achieving test R2 of 0.92 and 0.95, respectively. Acceptable predictive performance (test R2 of 0.82 and RMSE of 7.60) for the prediction of NH was also achieved. The feature importance analysis, partial dependence, and Shapely value were used to interpret the prediction models and study the NH formation mechanisms and behavior. Then, forward optimization of NH was implemented based on optimal predictive models, indicating the high potential of ML-aided bio-oil production and engineering.