Surface Tension Prediction of Fuel Additives Based on Machine Learning Model with Subtraction-Average-Based Optimizer Algorithm

Fuel additives play a significant role in improving combustion efficiency and fuel quality, with their surface tension being a crucial thermophysical property that directly affects atomization and cylinder performance. To address the demand for thermophysical data of fuel additives, 574 surface tension data for 22 fuel additives were extensively collected and evaluated using empirical models. A modified Sastri-Rao model (M-Sastri-Rao model) was built with critical temperature (T_c), reduced temperature (T_r), critical pressure (p_c), boiling point temperature (T_b), and acentric factor (ω) as influencing factors. The empirical models were found to have limited accuracy in predicting the surface tension. Then, a BP neural network model with the subtraction-average-based optimizer (SABO) algorithm was proposed. The results show that the SABO-BP model significantly reduced the deviation between calculated and experimental values, outperforming the previous empirical models. Various evaluation metrics were calculated for the SABO-BP model. The distribution of Bias ranged within ±5%, and the mean absolute error reached 0.165 mN·m^–1. The key parameters affecting the model were identified through a SHAP interpretability analysis. The SABO-BP model can accurately provide surface tension data for applications in the design and simulation.