Prediction of Biocrude Oil Yield From Biomass Hydrothermal Liquefaction Via Interpretable Machine Learning Using Higher Heating Value and Process Parameters

Abstract

The growing demand for sustainable energy calls for efficient and accurate methods to optimize biofuel production processes. Hydrothermal liquefaction (HTL) is a promising thermochemical technique to convert wet biomass into biocrude oil, but estimating yield across diverse feedstocks and conditions remains challenging. In this study, we develop and benchmark a series of machine learning models to predict biocrude oil yield from HTL, using a comprehensive dataset of 650 biomass samples and process parameters, including elemental composition and higher heating value (HHV). Notably, this is the first study to incorporate HHV as a predictive feature at this scale. Seven ML models-including XGBoost, Random Forest, and Gaussian Process Regressor-were optimized via Bayesian hyperparameter tuning and evaluated through a dual-validation strategy combining tenfold cross-validation with a hold-out test set. XGBoost achieved the highest performance (R2 = 0.97, RMSE = 0.033). To ensure model interpretability, SHAP and SAGE techniques were applied, identifying HHV, carbon content, and pressure as key yield predictors. These results provide a transparent, data-driven framework for enhancing reactor design and feedstock selection in bio-oil production systems. The study underscores the potential of interpretable ML in advancing the predictive capabilities of renewable fuel technologies.