Prediction of hydrocarbons ignition performances using machine learning modeling

This study presents a computational methodology for determining the Derived Cetane Number (DCN) of practical aviation fuels. The proposed approach integrates a novel Quantitative Structure-Property Relationship (QSPR) model designed to predict DCN for hydrocarbon species and mixtures with fuel composition analysis obtained through Two-Dimensional Gas Chromatography (GCxGC). The QSPR model used 20 independent variables computed based on selected hydrocarbon molecular descriptors, including functional groups and distance-based topological indexes. The multivariate regression analysis was used to train the QSPR model based on a dataset consisting of 48 individual hydrocarbon species and 157 surrogate mixtures. The model demonstrated robust predictive capabilities with high coefficients of determination (R2) of 0.96 on the training dataset and 0.94 on the independent testing dataset. The latter consisted of 43 surrogate mixtures formulated both in-house and sourced from archived literature. The application of the QSPR model for practical jet fuels involves specifying the detailed jet fuel compositions using GCxGC analysis and a randomization algorithm based on a database featuring over 17,000 distinct hydrocarbon species. The overall model's performance on practical jet fuels aligns closely with its performance on the training and testing datasets, affirming its practical utility. To enhance prediction accuracy of the proposed computational approach for practical jet fuels, density was explored as a potential constraining property to narrow randomization results with those detailed composition having a density similar to the actual fuel. In this regard, a novel relationship was established to predict fuel densities based on compositional characteristics. Despite the promising results in density prediction, this study indicates that density alone is insufficient to effectively constrain randomized compositions for significantly improved DCN predictions, and thus, further development is required.