Predicting FuelProperties through Sequential ForwardSelection (SFS) Enhance Ensemble Machine Learning with TopologicalIndex Selection

Predictive modeling of fuel properties is of paramount significance in fuel formulation. Topological index descriptors, sourced from the Mordred library and optimized through sequential forward selection (SFS), are employed to systematically develop models predicting the cetane number (CN), boiling point (BP), melting point (MP), and flash point (FP). Six models, extreme gradient boosting regressor (XGBR), gradient boosting regressor (GBR), linear regression (LR), random forest regressor (RFR), support vector regression (SVR), and voting regressor (VR), are assessed for efficacy across different sample sizes: 249 for CN, 701 for FP, 454 for MP, and 1054 for BP. The VR model, integrating XGBR, GBR, LR, RFR, and SVR, is demonstrated to be the most effective model overall. SFS identifies 16, 18, 16, and 15 key descriptors for CN, FP, MP, and BP, respectively. Shapley additive explanation (SHAP) plots highlight the critical descriptors for each property. A 10-fold cross-validation further confirms the VR model’s superior performance, exhibiting high accuracy across all data sets. This work presents a robust framework for feature selection and model development in the prediction of fuel properties.