Fuel property modeling by high-speed gas chromatography coupled with partial least squares data analysis

Abstract

Partial least squares (PLS) regression is a valuable chemometric tool for property prediction when coupled with gas chromatography (GC). Since the separation run time and stationary phase selection are crucial for effective PLS modeling, we study these GC parameters on the prediction of viscosity, density and hydrogen content for 50 aerospace fuels. Due to the diversity of compounds in the fuels (primarily alkanes, cycloalkanes, and aromatics), we explore both polar and non-polar stationary phase columns. The robustness for the PLS models was evaluated by their normalized root mean square error of cross-validation (NRMSECV). PLS models built for viscosity across 1-min, 3-min, 7-min, and 10-min time window (TW) high-speed GC separations produced nearly the same NRMSECV with the polar column data with an average (standard deviation) of 4.41 % (0.34 %) versus the non-polar column data of 4.69 % (0.15 %). In contrast, while the NRMSECV of density modeling with the polar column data varied more than the viscosity models, averaging 7.54 % (0.67 %), the non-polar column data produced a significantly higher average NRMSECV of 10.06 % (0.35 %). Similarly, for hydrogen content, the NRMSECV with the polar column data averaged 9.50 % (0.87 %), which was significantly lower than the NRMSECV with the non-polar column data averaging 12.10 % (0.88 %). We also investigated the impact of smoothing the GC data on the corresponding PLS models. By applying varying degrees of smoothing, we can effectively obtain similar chromatographic peak patterns in a shorter TW. For example, a 10-min smoothed chromatogram appears like the 1-min separation with no smoothing but resulted in nearly the same NRMSECV. Overall, the fast separation with a 1-min TW produced robust PLS models for viscosity with either stationary phase column, whereas for density and hydrogen content the polar stationary phase column produced superior PLS models, thus with proper stationary phase selection, a fast separation run time could be readily applied with optimal PLS property modeling results.