(267h) Prediction of Cetane Number Using Molecular Graph Modularity and Functional Groups in Artificial Neural Networks

Cristopher Arvizu Cano¹, Jorge Aburto², Elias Martinez-Hernandez²

¹ Faculty of Chemistry, National Autonomous University of Mexico

² Energy Efficiency and Biofuels Division, Mexican Institute of Petroleum

The cetane number is a measure of the ignition quality of a diesel fuel in a compression ignition engine and it is experimentally determined with an engine test standard (ASTM D613). A fuel with a high cetane number has a short ignition delay period and causes fast release and combustion of fuel after being injected into an engine where the air has been compressed and heated to a high temperature. A molecule is good candidate as an additive or fuel for diesel engines if it has a high number cetane. In this work, we developed a model to predict cetane number to identify candidate molecules derived from biomass by using molecular graph modularity and functional group counts as inputs to train an artificial neural network (ANN). This approach has been demonstrated as promising in physical property predictions of properties such as viscosity. A database from literature was collected and processed with the final data set containing 519 compounds of different class (alcohols, aldehydes, alkenes, aromatics, carboxylic acids, cyclo-alkanes, esters, ethers, furans, iso-alkanes, ketones, lactones, n-alkanes and polyfunctional), which were split into training and test sets. The ANNs were trained in Matlab using the ANN Toolbox with a single hidden layer and varying the number of hidden neurons to minimize the root mean squared error (RMSE). To test the benefit of including modularity, models consisting of only functional group counts were developed and compared. In addition, the prediction was compared against other existing models in the literature. After model reduction by removing functional groups with lowest relevance, the best model was the one incorporating modularity with a RMSE=5.61. Results of the RMSE showed that the new models decreased the prediction error by up to 45% or maintain the same level but with the advantage of not requiring other features such as calculated quantum chemical descriptors. The parity plots for each compound classes showed a correlation value >=0.9. In general, we found that molecules with higher modularity values and presence of long carbon chains, linear ester and ether groups lead to higher cetane number. Finally, we applied the model to screen potential derivatives from biomass finding candidates for centane enhancing additives.