Spectroscopic-Based DCN Prediction Using Functional Group Surrogate Informed Machine Learning Models
thesis
posted on 2023-12-01, 00:00authored byAnandvinod Dalmiya
Derived cetane number (DCN) is an important, standardized fuel property relating to ignition propensity in compression ignition engines. To enable future unmanned aircraft systems (UAS) to operate on a broader range of fuels, on-board fuel characterization, at least for DCN, will be extremely beneficial for ensuring ignition. Typical DCN measurement techniques like Ignition Quality Testers (IQT) or other standardized equipment, do not offer non-invasive, quick, and miniaturized pathways for onboard fuel sensing. The combination of an infrared (IR) spectral measurement of liquid fuels with robust Machine Learning (ML) models for DCN prediction, potentially opens pathways for onboard fuel sensing for engine control. Analysis of the different spectral ranges is performed in order to evaluate their capability to predict DCN with accuracy. Because spectral range and resolution trade off against size of sensors, finding optimal range is important and offers pathways to miniaturize the sensors. Existing models use similar types of fuels operating on a small and predefined range of DCN in order to achieve high prediction accuracy. These existing models fail to predict the DCN of fuel samples outside their training set. These models must be made robust enough to predict the DCN of new and untrained fuels within tolerable limits.
Absorption spectra of a few jet fuels, neat hydrocarbons, and their mixtures have been collected using Fourier Transform Infrared (FTIR) absorption spectroscopy. These spectra have been correlated to DCN for the purpose of predicting the DCN of real fuels and their blends. A method has been developed to associate DCN using ML models by avoiding saturated regions in these spectra. Several spectral ranges including the full-range (4000-800cm-1), extended fingerprint region (2400-800cm-1), and windows of narrow IR regions (ranges <300cm-1) have been investigated to evaluate their importance in DCN prediction and prediction accuracy. In order to expand the predictive capability of fuels not present in the training set, the use of neat hydrocarbons and their mixtures representing chemical functional groups present in real jet fuels is evaluated for constructing a calibration model predicting real fuels operating on a broader range of DCN.
Additionally, prediction performance between FTIR (transmission-based) and Attenuated Total Reflectance (FT-ATR) spectroscopy is compared to evaluate the impact of saturated regions on DCN estimation. In transmission-based absorption, due to the difficulties in fabricating a narrow pathlength (<10µm) liquid cell, the strong absorption bands located near 2900cm-1 in the mid-IR region experience saturation. Our methods can exclude these saturated regions in correlation to DCN. Alternatively, ATR spectroscopy has the advantage of acquiring spectra without any saturations due to its typical pathlength of ~2µm in most of the liquid samples. After comparison, the results indicate the importance of saturated regions in correlating IR spectra to DCN.
Finally, elucidation of chemical functional groups from IR spectra is demonstrated to estimate the composition of functional groups present in the sample. This method is evaluated as one means of expanding the capabilities of the ML models to predict new and unseen fuel beyond their training dataset, that is by comprehensively understanding the relationships between spectra and a representation of chemical functional groups and the relationships between chemical functional groups and DCN separately. Such a method could be used to correlate to other physiochemical properties of fuels expanding the capabilities to better enabling the engine control. In this process, uncorrected FT-ATR spectra of pure hydrocarbons samples and their mixtures spanning the range of functional groups found in the real fuels are used to train the spectral model to estimate the composition of chemical functional groups present in the real jet fuels. Subsequently, the predicted composition is fed to a functional group-based ignition model, predicting the DCN of the fuel. Ultimately, this novel method shows fair to good accuracy. It is compared against other methods, and suggestions for future enhancements to the models are discussed.
History
Advisor
Patrick T Lynch
Department
Mechanical and Industrial Engineering
Degree Grantor
University of Illinois Chicago
Degree Level
Doctoral
Degree name
PhD, Doctor of Philosophy
Committee Member
Kenneth Brezinsky
Hadis Anahideh
Igor Paprotny
Scott T. Sanders