NIR With Problem Data Sets

There are many approaches to developing a data set for Near Infrared calibrations, some statistically valid, some not; but all with some interesting insights. There is the ideal situation with a large set of samples, ample laboratory resources for reference chemistry. Do you make a random selection or take every nth sample and scan and analyze? Do you start scanning and analyzing and build your calibration data set gradually as you go, with the big question? When do I stop? Yes, you can use the reference data to select a diverse range of samples insuring that the ends as well as the range are adequately sampled. Some chemometric software packages will choose the most diverse set of samples based on the spectra. This provides a stopping point for initial calibrations and limits the amount of reference chemistry you may have to do. However, it is often not the situations that arise in the NIR lab that provide the most strenuous challenges. Usually it is the real world samples from commercial operations that have their limitations based on the process. The bottom line is you get to work with what you are handed. We will examine some of these data sets and the unique opportunities for model development.