Raman Spectroscopy-Driven Chemical Functional Group Modeling for Fuel Property Prediction

Rapid, low-volume
assessment of ignition quality and
density of
aviation fuels is increasingly important as more synthetic fuels are
integrated into the fuel supply. We present a two-tier chemometric
framework that couples Raman spectroscopy with functional-group-aware
machine learning to predict derived cetane number (DCN) and liquid
density. A curated set of 160 liquids, including neat hydrocarbons,
their mixtures, representative jet fuels and jet fuel mixtures, was
measured with a compact Raman system. Literature-guided peak assignments
were consolidated into a functional-group table mapping Raman bands
to “spectroscopic” groups, and translated into 11 hydrocarbon
UNIFAC descriptors. Tier-1 uses a multitarget artificial neural network
(ANN) to regress selected Raman channels (632 features) onto the UNIFAC
composition vector. Tier-2 uses the predicted UNIFAC vectors to estimate
bulk properties using an ANN for DCN and linear regression for density.
Group-aware data splitting and cross-validated hyperparameter tuning
were applied uniformly. Tier-1 reproduces functional-group compositions
with high fidelity (test-set R2 ≥
0.95 for all UNIFAC groups). The DCN model attains the test, R2 ≈ 0.94, with ∼74% of samples
within ±10% of experimental measurements. Further, a Monte Carlo
sensitivity analysis reduces R2 by only
a few points, indicating robustness to spectral noise. For density,
a multivariate linear regression suffices (test-set R2 = 0.90, with 91% of samples within ±2% of experimental
measurements). Subsequent permutation-importance analysis aligned
with known structure–property trends. Together, these results
demonstrate a fast, interpretable route from Raman spectra to composition-based
property predictions suitable for real-fuel screening and, ultimately,
deployable sensing.