Smoke point prediction of oxygenated fuels using neural networks

Abstract

Smoke point (SP) is an important fuel property that characterizes the propensity of aviation jet fuels and kerosene to form soot. In the present study, an artificial neural network (ANN) model based on the artificial intelligence principle was developed to predict the SP of fuels with oxygenates (ethers and alcohols) and hydrocarbons (e.g., paraffins, olefins, naphthenes, aromatics, and their blends). An experimental dataset of 366 fuel mixtures comprising 113 pure compounds, 8 types of gasoline and diesels, and 245 fuel blends was used to improve the SP-prediction performance of the ANN model. One hundred and ten of the fuel blends used were various jet fuel surrogates collected from the literature. Experimental SP of 40 new mixtures consisting of diethyl ether, dibutyl ether, dioctyl ether, diphenyl ether, and methyl tert-butyl ether along with gasoline and diesel was measured as part of this work. The molecular composition of the fuels was expressed as the weight % of eight constituent functional groups present in the fuel. These functional groups along with two additional parameters (branching index and molecular weight) were provided as ten inputs to the model. The functional groups present in the gasoline and diesel samples were determined using high-resolution 1H nuclear magnetic resonance spectroscopy. The data were randomly split into three sets: training (70 %), validation (15 %), and test (15 %). The model was initially trained and validated simultaneously, and then; it was tested. A positive linear relation was observed between the measured and predicted SPs, as indicated by a correlation coefficient of 0.98. The mean absolute error of the predicted SP was 4.5. Results showed that the SP of the fuels depended on the aforementioned parameters that served as inputs for the model. The proposed model can be used to predict the SP of pure and blended form fuels containing the aforementioned functional groups.