Four supervised regression-based machine learning (ML) models, including random forest (RF), support vector regression (SVR), light gradient boosting regression (LGBR), and extreme gradient boosting regression (XGBR), were employed for predicting catalytic oxidative desulfurization (ODS) of liquid fossil fuels using ionic liquids (ILs) as catalysts and H2O2 as the oxidant. The models were developed using 1383 experimental data points derived from ODS utilizing 31 various ILs. Six input variables, including process temperature in the range of 25–100 °C, reaction time in the range of 5–360 min, oxidant-to-sulfur (O/S) molar ratio in the range of 2–100, IL type, feedstock type, and extractant type, and one output variable of sulfur removal (%) were considered. Among these models, XGBR indicated the best performance, achieving the lowest root-mean-square error (RMSE = 2.72), mean absolute error (MAE = 1.63), and mean absolute percentage error (MAPE = 4%), along with the highest determination coefficient (R2= 0.99).
Boshagh et al. (Sun,) studied this question.