DOI: 10.1021/acs.energyfuels.8b00556
Contribution
In this work, I focused on harnessing Artificial Neural Networks (ANNs) to accurately predict the Research Octane Number (RON) and Motor Octane Number (MON) for various fuel blends, a significant step forward in understanding complex fuel behaviors. My work centered around developing and fine-tuning the ANN models, integrating molecular parameters obtained from 1H Nuclear Magnetic Resonance (NMR) spectroscopy data. This integration was crucial for capturing the intricate relationships between molecular compositions of fuels and their octane ratings, especially in the context of non-linear variations observed with ethanol-blended gasoline.
My contribution extended beyond the model’s initial development to include a rigorous validation process. I ensured that the ANN models for RON and MON, which comprised two hidden layers and a large number of nodes, were accurately trained to reflect the complex interplay of molecular features affecting fuel performance. The validation process involved comparing the ANN’s predictions with a separate test set of experimentally measured RON and MON values. This comprehensive approach not only demonstrated the effectiveness of the model but also highlighted the potential of machine learning tools like ANNs in capturing complex chemical relationships, a foundational aspect that I would build upon in my subsequent research works.
Abstract
Machine learning algorithms are attracting significant interest for predicting complex chemical phenomenon. In this work, a model to predict research octane number (RON) and motor octane number (MON) of pure hydrocarbons, hydrocarbon-ethanol blends, and gasoline–ethanol blends has been developed using artificial neural networks (ANNs) and molecular parameters from 1H nuclear magnetic resonance (NMR) spectroscopy. RON and MON of 128 pure hydrocarbons, 123 hydrocarbon–ethanol blends of known composition, and 30 FACE (fuels for advanced combustion engines) gasoline–ethanol blends were utilized as a data set to develop the ANN model. The effect of weight percent of seven functional groups including paraffinic CH3 groups, paraffinic CH2 groups, paraffinic CH groups, olefinic −CH═CH2 groups, naphthenic CH–CH2 groups, aromatic C–CH groups, and ethanolic OH groups on RON and MON was studied. The effect of branching (i.e., methyl substitution), denoted by a parameter termed as branching index (BI), and molecular weight (MW) were included as inputs along with the seven functional groups to predict RON and MON. The developed ANN models for RON (9-540-314-1) and MON (9-340-603-1) have two hidden layers and a large number of nodes, and was validated against experimentally measured RON and MON of pure hydrocarbons, hydrocarbon–ethanol, and gasoline–ethanol blends; a good correlation (R2 = 0.99) between the predicted and the experimental data was obtained. The average error of prediction for both RON and MON was found to be 1.2 which is close to the range of experimental uncertainty. This shows that the functional groups in a molecule or fuel can be used to predict its ON, and the complex relationship between them can be captured by tools such as ANNs.