Machine Learning-Based Prediction of Sustainable Aviation Fuel Yield using Literature-Derived Hydroprocessing Data

Abstract

The increasing demand for sustainable aviation fuel (SAF) has encouraged the development of efficient predictive approaches for optimizing jet fuel production from renewable feedstocks. Conventional experimental optimization methods are often time-consuming and expensive because hydroprocessing performance is strongly influenced by feedstock characteristics, catalyst composition, and operating conditions. In this study, machine learning (ML) techniques were applied to predict jet fuel yield using a dataset compiled from approximately 50 published scientific articles. The dataset consisted of 101 experimental observations involving different feedstock groups, catalyst metal groups, catalyst supports, catalyst loading, free fatty acid (FFA) content, temperature, pressure, and weight hourly space velocity (WHSV). The ML workflow was developed using Orange Data Mining software and included data preprocessing, feature selection, imputation, model training, and performance evaluation. Four regression algorithms, namely Random Forest, Linear Regression, Neural Network, and Gradient Boosting, were evaluated using 10-fold cross-validation. The Gradient Boosting model achieved the best predictive performance with an RMSE of 7.172, MAE of 5.314, MAPE of 10.026%, and R2 value of 0.286 during cross-validation. Feature ranking analysis indicated that catalyst support type, feedstock group, catalyst metal group, and FFA content were among the most influential variables affecting jet fuel yield.