Strategy for constructing robust multivariate calibration models

In multivariate calibrations usually a minimal residual error in the modelʹs predictions is aimed at, while generally less attention is paid to the robustness of the model with respect to changes in instrumentation, laboratory conditions, or sample composition. In this paper, we propose a strategy for selecting a multivariate calibration model which possesses a small prediction error and, simultaneously, is less sensitive to the above-mentioned variations. The strategy is applied to calibration models used to predict the density of poly(ethylene terephthalate) (PET) yarns from the Raman spectra. The strategy implies that spectra of calibration samples are measured under circumstances under which the application will be implemented, and spectra of a smaller set under different conditions (variations in ambient temperature, laser power, and laser frequency) according to an experimental design. The prediction results of the calibration model are used in a ruggedness test in order to test the sensitivity. In this study various calibration models using different spectral preprocessing techniques are tested. These ruggedness results together with the prediction error are used to select a good model. Moreover, it is possible in this way to provide the boundaries for the experimental conditions, where the model is valid.