Raman Chemometric Method for Fuel Property Prediction

Multi-fuel compression-ignition engines for unmanned aircraft systems require real-time knowledge of fuel ignition quality and related properties. Conventional measurements of derived cetane number (DCN) and density are accurate yet slow, sample intensive, and confined to laboratory instruments. This thesis develops and validates Raman-based chemometric methods that enable rapid, portable estimation of these properties and demonstrates their feasibility in online and engine-integrated settings. A curated database of 168 liquid-phase samples was assembled that spans the functional-group landscape and property ranges of standard jet fuels. The set includes neat hydrocarbons, laboratory mixtures, Army Research Laboratory CN surrogates, actual jet fuels, and fuel blends. Raman spectra were preprocessed with Savitzky-Golay second-derivative filtering and Standard Normal Variate normalization to suppress baseline drift and multiplicative effects. Group-aware data splits kept replicate scans within a single fold to avoid leakage and to produce honest estimates of generalization. Two complementary modeling approaches were pursued. The first is a direct mapping from spectra to DCN using supervised machine learning (ML) models. Out of the many ML approaches investigated on three different subsets of the Raman spectral range, artificial neural network (ANN) delivered a strong test performance on the fingerprint region with more than ninety-six percent of unseen test samples within ten percent experimental error. A global surrogate analysis identified one hundred fingerprint-region features influencing DCN predictions which supported a reduced ANN that improved accuracy while reducing the input dimension. Two prototype studies (Gen-1 and Gen-2) translated these models to practice. The Gen-1 Raman sensor with a flow cell and a Raspberry Pi computer produced continuous predictions about every seventeen seconds and maintained accuracy through controlled temperature settings. On the other hand, an engine test-cell demonstration integrated the Gen-2 Raman sensor in line with the fuel system and tracked live fuel switches between low and high ignition quality blends. Predictions remained within prescribed error bands and agreed with laboratory Ignition Quality Tester measurements. Together these results establish a pathway to miniaturized Raman sensors that deliver fast, chemically grounded estimates of ignition quality for real-time control and fuel screening. Finally, the second modeling approach introduces a two-tier framework that first predicts UNIFAC functional-group weight fractions from chemically-informed Raman features mapped using a curated functional group table and then maps those compositions to bulk properties. The Tier-1 multi-target ANN predicted eleven UNIFAC group weight fractions with high fidelity. The Tier-2 regressors then predicted DCN with test R2 near 0.94 and a median absolute error near three DCN units, and predicted density with near-linear behavior using a simple linear model with test R2 around 0.90 and mean absolute error near 0.014 g cm-3. Monte Carlo experiments that injected realistic composition noise showed modest degradation and confirmed robustness and Permutation Importance analysis revealed that the methyl (CH2) group dominates predictive power.

Author dc:creator

• Dhananjay Ambre (23589795)

• Raman spectroscopy
• Chemometrics

Statement dc:rights

• In Copyright

DOI dc:identifier

https://doi.org/10.25417/uic.32991839.v1

OAI identifier oai:identifier

oai:figshare.com:article/32991839