Thermophysical properties of used frying oil biodiesels blended with alcohols: Robust machine learning frameworks for density prediction

Biodiesel–alcohol mixtures from waste frying oil are emerging as renewable, engine-compatible fuel candidates, yet their thermophysical characterization remains incomplete. In particular, density is a critical property for process and engine modeling, but experimental datasets are fragmented and predictive frameworks scarce. In this study, a comprehensive dataset of 2389 density measurements was assembled for used frying oil biodiesel (UFOB) blended with different alcohols, spanning broad ranges of conditions. Three machine learning models were developed and evaluated: Gaussian Process Modeling (GPM), Adaptive Neuro-Fuzzy Inference System (ANFIS), and Decision Tree (DT). Among them, GPM achieved the highest predictive accuracy, with a coefficient of determination (R 2 ) of 99.68 % and a mean absolute percentage error (MAPE) of 0.17 %. ANFIS and DT also performed well, with MAPE values of 0.25 % and 0.72 %, respectively. Validation through scatter and distribution analyses, together with 5-fold cross-validation, confirmed the robustness and generalizability of the models. Beyond accuracy, the models were able to reproduce consistent physical behaviors of biodiesel–alcohol systems, reinforcing their reliability as surrogates for experimental studies. Model interpretability was further addressed using Shapley Additive Explanations (SHAP). Both global and local SHAP analyses consistently highlighted reduced temperature and composition of alcohol as the most effective factors governing density, followed by pressure and structural descriptors of the biodiesel. These outcomes demonstrate that combining predictive accuracy with physical consistency and explainability offers a reliable pathway for modeling UFOB–alcohol blend densities, thereby supporting improved process design and renewable fuel development. • A total of 2389 experimental density data points from literature were analyzed. • The GPM, DT, and ANFIS algorithms were applied for accurate density prediction. • The GPM model showed the best performance with R 2 = 99.68 % and MAPE = 0.17 %. • Predicted density trends closely matched experimental observations. • SHAP analysis revealed temperature as the most impactful factor.