Using Machine Learning to Predict Derived Cetane Number and Fuel Similarity

Abstract

Abstract Nearly four hundred different samples of jet and diesel fuels were used to train and test Machine Learning (ML) models for Derived Cetane Number (DCN – ASTM D6890) prediction using eight of the fuels’ physical properties as model inputs. Linear Regression (LR), Artificial Neural Networks (ANNs) and Gaussian based models all showed good performance predicting DCN with nominal prediction errors of 1 to 1.7 cetane numbers (CN). Shallow ANNs showed comparable prediction results as compared to LR, with the Gaussian Exponential Model yielding the best results overall. The DCN prediction models were exercised to observe the most critical-sensitive properties in the DCN prediction. Fuel density and T50 were seen to be the most important for both jet and diesel fuels. This result supports the usage of these two properties in cetane number prediction via the Cetane Index (CI) calculation (ASTM D976). Flash point and Tend of the distillation curve were of secondary importance. Additionally, jet fuel chemical composition data from 8 chemical fuel classes were applied to predict DCN. Adding the chemical composition data to the physical property data did not provide for improved DCN prediction. This result supports the coupling and connection between a fuel’s physical and chemical properties. An analysis of the most important (to DCN) fuel classes shows alkanes (high cetane) and alkyl-benzene (low cetane) components to be the most influential. Finally, fuel similarity was characterized using Self Organizing Maps (SOMs). The SOM map was trained for both jet and diesel fuels using physical properties alone. Different fuels (e.g. alternative Alcohol-to-Jet) were then applied to the SOM to test similarity. SOM Position and Quantization Error are shown to accurately characterize these fuels as significantly different than the conventional jet and diesel fuels used to establish the SOM.