| Author | Message | ||
     Licen Ziaato |
I am trying to use NIR to measure the amount of different components in a blend. Is it possible to create a virtual spectrum using the spectra of all the used components ? Is there any algorithm to do this or am I dreaming ? Rgds Licen | ||
| hlmark |
Licen - In an ideal world this would work. In the real world, however, procedures like that tend to fail when the resulting model is used to analyze scamples from an actual production process. The reason is that the virtual samples do not take into account the extranous variabilty that the process imposes on the samples. These include the variations due to changing process conditions (temperatures, differences and changes in starting materials, mechanical effects, etc.) These are manifested through the variations in optical effects that are measured. Even without those, there are known effects due to variations in optical scatter that are present, at least in solid samples, and also in liquids unless they are completely clear. There are also effects of noise, non-linearity, interactions between sample components, etc., that would not be modeled by the virtual model. Those are the reasons why calibrations should be based on real production samples. Sometimes, virtual samples or even better, laboratory samples or pilot-plant samples can be used to extend the range of sample variability when the production samples do not have sufficient natural variation. However, caution must be used that the behavior characteristics of the non-production samples do not overwhelm the production samples. If possible, the laboratory samples should show as much of the same types of variabilities as actual production samples do. Howard \o/ /_\ | ||
| Vitas Svedas |
Licen, you are not dreaming. In some cases it is possible to relate (to deconvolute) spectrum of blend to spectra of chemical components. Also, generation of adequate virtual spectrum is attainable by the same way. Recent successful application of such deconvolution to chemical components is published by Harald Martens et.al. in Anal.Chem. vol.75, p. 394 (2003). The 0.5 % accuracy of predicted concentration of wheat gluten and wheat starch mixtures was achieved. Unfortunately, general suitability of such methods to the real analysis is limited. This is because of complexity of real samples, their intricate interaction with the radiation and of the processes inside the radiation detectors. These and the other reasons were disclosed here by Howard. Probably models like published in Anal.Chem. will step-by-step conform to more complex samples and to the radiation measurement situation. In that case the round of deconvolution-treated spectra will grow. Vitas | ||
| hlmark |
Yes, what Vitas said about Harald's work is correct, but Harald used only a set of two-component mixtures, and even then he had to use extensive preprocessing of the spectra in order to obtain optimum results. In fact, the whole point of the paper was that ordinary MSC as a preprocessing method did not suffice, and he had to use an extended variation of that data pretreatment method in order to account for the effects of optical scatter. \o/ /_\ | ||
| Gabi levin |
Dear Licen, One of the first things I was taught, and also learned the hard way in an instance or two was that if I want to do a feasibility I should try to obtain samples from a real process, or product, rather than try to rely on "synthetic" samples whereby I would mix ingredients in known proportions. The reason for preferring real samples is that they bring into the game all the variances in contaminants, particle size distribution, etc., that real life application will bring. Of course, many times "synthetic" samples are useful for preliminary evaluation of applicability of NIR, etc. Still, the ultimate judge of the possible real life success is sampling and representing the process. To take the synthetis one step further and artificially create spectrum from the individual spectrum of each component is a larger risk than creating "synthetic" samples. I am not absolutley sure the benefits of doing so will outweigh the possible risk that the entire effort will have to be dumped down the drain. One more thing, while in transparent liquids this approach may have some more merit, (granted that there are no inetractions between the chemicals, such as hydrogen bonding, dipole interactions, etc. that will affect the "basic" spectra of the components) in powders the situation is worse because bewteen the time you collect the "pure" spectra of the individual components, to the time that you collect the spectra of the mixtures you process the powders. The processing is the mixing. In practice during the mixing the particle size can change, thus introducing changes that can not be accounted for by the "synthetic" spectrum. Sorry for being skeptic, I am a strong believer in hard work, and getting the fingers dirty. Gabi |