An assessment of multivariate predictions of various properties of crude oil on two ATR/IR spectrometers

Abstract

About sixty commercially relevant crude oil samples are acquired on two spectrometers that are linked by a calibration transfer function and analyzed through PLS regression. About twenty multivariate models for distillation fractions are built and tested, as well as the models for API, acids, and sulfur. The parallel acquisition of samples on both spectrometers allows for deriving calibration and predictions errors for the models that use calibration transfer function as well as for the models independently built and tested on each of the used spectrometers. Testing of all available samples reveals that the calibration transfer approach produces notably higher prediction errors in comparison to the testing on the spectrometer on which the models are developed. Splitting the samples so that exact same sets for calibration and testing are used on both spectrometers (no calibration transfer) reveals that one spectrometer performs clearly better than the other although both operate under the same operating procedure and pass the same operational qualification. Comparison of the spectra of crude oils acquired on both spectrometers that are consistently identified as outliers on one of the spectrometers does not indicate any easily identifiable spectral differences. The results of this study indicate that the prediction errors may significantly differ depending on the performance of the spectrometers employed despite them nominally passing the qualification criteria, thus being nearly equivalent.