Toward biodiesel-additive modeling: Explainable intelligent frameworks for thermal conductivity prediction

Accurate thermal conductivity estimation is critical for optimizing fuel atomization and combustion efficiency in modern engines. This study introduces a high-fidelity data-driven framework for predicting the thermal conductivity of biodiesel–additive blends across extensive thermodynamic ranges, including 249–523 K and pressures up to 42.75 MPa. Utilizing a comprehensive repository of 2811 experimental data points, four computational paradigms, including least-squares boosting (LSBT), radial basis function neural networks (RBFNN), adaptive neuro-fuzzy inference systems (ANFIS), and bagged trees (BT), were developed using a unified seven-dimensional input vector. Bayesian optimization was employed for hyperparameter tuning to ensure a robust bias–variance trade-off. Statistical evaluations revealed that the LSBT model provided the best predictive performance, achieving a validation mean absolute percentage error (MAPE) of 0.56%, a relative root mean squared error (RRMSE) of 0.72%, and an R 2 of 99.52%. Robustness was verified through 5-fold cross-validation, exhibiting a negligible generalization gap, while William’s plot analysis confirmed that 99.08% of the data resided within the valid applicability domain. Notably, the developed models reproduced physically consistent thermal conductivity trends, particularly the non-monotonic relationship with molecular chain length and an apparent transition around 8 carbons, which may reflect changes in molecular packing, polarity, and dispersion-dominated interactions as chain length increases. Furthermore, SHAP-based interpretability analysis identified temperature as the most critical driver while distinguishing between associative alcohol effects and van der Waals interactions in alkanes. This optimized framework serves as an accurate and interpretable data-driven surrogate for rapid thermophysical property estimation, reducing experimental effort and supporting large-scale computational tasks and industrial combustion modeling within the investigated domain.