Interpretable machine learning model for bio-oil property prediction from hydrothermal liquefaction of biomass via graph neural networks-based molecular structure

Abstract

Hydrothermal bio-oil is a promising fuel source to effectively alleviate the energy crisis and contribute to carbon neutrality. However, the application performance of hydrothermal bio-oil is primarily affected by its yield, compositions (C, H, O, N, and S contents), and higher heating value (HHV). Hence, we proposed an interpretable machine learning model for hydrothermal bio-oil property prediction. Graph neural network (GNN)-based molecular structure was utilized to enhance datasets. Single- and multi-task prediction models were established by extreme gradient boosting (XGB), random forest (RF), gradient boosting decision tree (GBDT), and deep neural network (DNN). Shapely values and partial dependence were utilized to interpret the impact of input features on output responses. Results indicated that XGB model (average train R² = 0.95, RMSE = 1.29, MAE = 0.69, average test R² = 0.91, RMSE = 1.30, MAE = 0.93) performed the most optimally after applying GNN, with an average performance improvement of 6.74–7.95 %. The model demonstrated outstanding performance on unknown data, achieving an average R² of 0.916. Temperature and biomass ultimate analysis emerged as pivotal features influencing output. Triglycerides exhibited a stronger influence than fatty acids owing to their higher carbon content. A combination of high temperature (>350 °C) and elevated triglyceride content (>30 %) increased bio-oil yield (∼40 wt%) and HHV (∼35 MJ/kg).