Emerging opportunities using different calibration approaches

This work reviews different calibration methods used in multivariate calibration. A common feature in these methods is that they aim at using a priori information in calibration. Some of their potential has not been fully recognized and used. We demonstrate these with different approaches. In the first approach, both generalized least squares (GLS) and partial least squares (PLS) will be used in a challenging calibration problem with only three pure component samples and one mixture sample. To further illustrate the full potential of these techniques, we also provide an example where we do not measure the pure component spectra but estimate them with using the independent component analysis (ICA) and a few mixture samples without reference lab data. The above approaches will be tested using a multipoint NIR instrument with two 5-channel measurement probes. These models will be developed using data from only one channel and other channels and the second measurement probe will be used as an independent test set. The second approach illustrates the importance and advantages of regularization. Relatively new methods such as least absolute selection and shrinkage operator (LASSO) and elastic nets provide interesting opportunities for model development and robustification. This point will also be exemplified using the multipoint NIR data. In the third approach, we test some new techniques to cope with different nonlinear absorption-band related interactions. The above multipoint data set will be used here. Interesting point is that the data set contains three chemical constituents (ibuprofen, lactose and MCC) and there are 17 mixtures "replicated" three times but with different levels of mean particle size of lactose. In summary, this work addresses new and interesting directions in developing calibration models in the field of spectroscopy, multipoint measurements, and chemical imaging. Though we seem to favor the use of a few samples instead of many, we fully recognize and wish to point out that in that case the quality of data becomes even more important, and also that the model validation should be based on large enough, representative and independent validation and test sets. In this respect, it would be extremely unwise to rely on only a few validation and test samples.