In the aviation industry, sustainable biofuels are emerging as a crucial alternative to reduce dependence on fossil fuels and mitigate harmful greenhouse gas emissions. However, determining the physicochemical properties of biofuel blends traditionally relies on expensive and time-consuming laboratory experiments. This study proposes a highly accurate, data-driven computational approach to predict the density of biofuel blends obtained by mixing 14 different plant and animal-based oils with JP-5 jet fuel at various ratios. To ensure robust generalization and eliminate overfitting risks on the experimental dataset (71 samples), six advanced machine learning architectures—Artificial Neural Networks (ANN), Gaussian Process Regression (GPR), Regression Trees, Random Forest, LSBoost, and Support Vector Machines (SVM)—were comprehensively evaluated using a rigorous 5-fold cross-validation strategy. The results demonstrated that the Artificial Neural Network optimized with the Bayesian Regularization algorithm (ANN-BR) achieved the highest predictive performance. Specifically, the ANN-BR model yielded a Cross-Validation Coefficient of Determination (R2) of 0.9820, a Correlation Coefficient (R) of 0.9910, and a minimal Mean Squared Error (MSE) of 0.00121 on the unseen test folds. The Regression Tree and GPR models also exhibited exceptional accuracy, closely following the ANN. Ultimately, this study proves that predictive machine learning modeling can reliably supplement and accelerate conventional fuel characterization tests, offering significant time and cost advantages for the aviation sector.
Bülent Kurt (Thu,) studied this question.