| Author | Message | ||
     Ralph Schneider |
Hello together, has anyone ever tried to perform a deconvolution on NIR-spectra of mixtures to resolve the pure spectra of each component? I thank you for your help. Ralph | ||
| campbell |
Ralph, There have been many efforts made, but in my estimation, little to show for it. The difficulty is we don't really understand all the interactions that take place between the various constituents, such as solvation in liquids, scatter in solids, temperature effects, etc. I know there are many that have tried, one place being Bruce Kowalski's group comes first to mind, but there are others. There has been some success but, again in my estimation, not nearly enough to replace the more "traditional" methods of MLR, PCRA and PLS. It would be nice to be able to use first principles for calibrations to perform both qualitative and qualitative analyses, but it doesn't appear to be in the cards soon. Don't take this as a strict negative, as I think continued effort will enable progress to be made. I just don't know how far the progress will go. Bruce Campbell | ||
| hlmark |
Ralph - To some extent, what Bruce says is true enough, but there are still things that can be done. This topic was addressed a long time ago, and you might want to read the following paper: Honigs, D.F., Hieftje, G.M. and Hirschfeld, T; "A New Method for Obtaining Individual Component Spectra from Those of Complex Mixtures"; Applied Spectroscopy, 38(3); p.317-322 (1984). As you can see, the "new" method is quite old at this point. However, some of the general points discussed are as valid today as they were then. In particular, the spectral reconstruction algorithm described is compared to quantitative analysis algorithms, and the point is made that the properties of the data that make for good spectral reconstruction are the same as those that make for good (i.e., accurate) calibration performance. Once you grasp the concept behind the algorithm that Dave Honigs describes, it becomes clear how it can be extended to use Principal Componennts or other constructs to perform the same task, albeit with the same caveats. As I recall, there were some other papers published around the same time ("same" meaning +/- five years or so) where other reconstruction algorithms were described, although I don't have any specific references. A good literature search should turn them up, though. I'm pretty sure that at least some of them were also in Appl Spect. Howard Mark | ||
| Tony Davies (Td) |
PCA is probably the most direct method see paper by Bertrand A search of the NIR Publications CD-ROM gave five possible papers. Two from Proceedings of NIR-99 and three from JNIRS: NIR-99 313-6 Evaluation of reconstruction algorithm efficiency by using image quality parameters R. Esposito, P.L. Indovina, A. Lauria and M. Lepore Istituto Nazionale di Fisica della Materia, Unità di Napoli, Dipartimento di Scienze Fisiche, Università �Federico II�, Napoli, Italy. NIR-99 483-491 Interfacing fast spectroscopy and molecular biology reveals mechanisms of light reactions in photosynthesis Gianfelice Cinque,a Roberta Croce a,b and Roberto Bassi a a Facoltà di Scienze, Università degli Studi di Verona, Strada Le Grazie, I-37134 Verona, Italy. b Max Plank Institut für Strahlenchemie, Mulheim am der Ruhr, Stilftstrasse 34, Germany. JNIRS JNIRS 4,75-8 (1996) Bertrand at el PCA JNIRS 7, 229-240 (1999) Wesley et al. JNIRS 5,167-173 (1997) Remote sensing people. | ||
| DJDahm |
While the iteractions mentioned in Bruce's message are sources of problems in deconvolution, I would like to mention (yet again) my favorite candiate for all the ills that befall man-kind; the fact that the spectra we we use are generally non-linear funtions of absorption. Because of this non-linearity, the spectral contibution of a component at low concentration will have a different shape than it will have at high concentration. Further, the supression of the components contribution at high absorption will be a function of the total absorption at a wavelength, meaning that the shape of the deconvoluted spectra will be depenendent on the total compostion of the sample. Add that to the facts that the spectral shape of particulate samples is dependent on particle size and void fraction, and some of the problems encountered in deconvolution become expected. My contribution in the "continued effort" is looking for ways to "linearize" the spectra. It is my personal belief that various forms of Principal Component Analysis will beifit more from the linearization of spectra that the PLS and MLR analyses for quantitative analyses. Don Dahm | ||
| David Russell (Russell) |
One of my colleagues once tried classical least squares to resolve a simple liquid mixture matrix. He had the benefit of knowing exactly what he had since the mixtures were made in his lab. However, even in this "ideal" case he met with limited success due to matrix effects. | ||
| hlmark |
Don (Dahm) makes some good points, but even that coin has two sides. One problem with attempts to interpret reflectance spectra even in the early days was/is the recognition that in complicated matrices, the spectra of the constituents differ from that the spectra of those material in the pure form. This is due to real spectral changes due to interactions of the material with the matrix. One of the things that "spectral reconstruction" algorithms could do for you, at least in principle, is to allow you to ascertain the spectrum of the material in the matrix, with all the real effects acting on it. This would be an enormous boon to theoreticians and experimentalists who would like to figure out the real chemistry going on in sample sets. "All" you need to be able to do is to trust the spectra you compute! Don, we're counting on you!!! Howard | ||
| td |
Do we have any agreeed definition of "matrix effects (ME)"? I think of ME as being about physics - the scattering of energy by the sample. What David is talking about is chemistry. Probably hydrogen bonding. The important thing to remember is that NIR measures what is really there not what some chemist "knows" he made up! I want to say something about Don's comment but I need some thinking time! Tony | ||
| DJDahm |
I had a recent conversation with Karl Norris. He said that he wondered whether the sample set that Dave Russell was taking about was liquid or powder. (Tony seemed to assume that it was a liquid.) Presumably a liquid in transmission with a small area detector will give "linear" spectra as far as the effects that I talked about earlier are concerned (until the absorption is so staong that the effects of stray light appear). Then the effects that are observable (that Bruce mentioned and that Tony calls "Chemistry") should be the only thing going on. In the model systems that Karl and I have been using, the polyethlene spectum of a single sheet had a different transmission spectum than a partial momolayer of particulate polyethylene. Thats not surpisinmg, since they were from different sources, but it made it harder to study the "diffraction physics" of the system, because of this "chemical effect". Karl then challenged me to think more broadly about the consequencses of the fact that we are obtaining absorption by observing changes in scatter. There are different consequences in this case than in the case of obtaining an absorption spectum by observing the loss of direct transmision. (He cited some data that Fred McClure had shared with him.) I believe that we can begin to understand the total if we isolate the peices and study the effects more individually. Anybody have any thoughts about some model systems that we shouod be looking at to light the way in understanding what the "real spectra" of our samples are? Don Dahm | ||
| hlmark |
There are at least two kinds of "matrix effects" that can distort NIR spectra. One type is the one alluded to above: the interactions between different materials in a mixture. The classic example is water-alcohol mixtures: mixing together one liter of each gives <2 liters of mixture. In the face of that, who would want to argue that the spectra are unaffected? Another "matrix effect" is not as complicated, but may be somewhat unappreciated. If you mix together water and actone, the absorbance at the wavelngths corresponding to the peaks of the -CH absorbance bands in the acetone DECREASES as the amount of actone INCREASES. "How can this be?" you ask. Very simple: sonce the water absorbance spectrum completely overlaps that of the acetone, and the absorptivity of the water is much higher than that of the acetone -CH bands, the displacement of the water decreases its absorbance faster than the increased concentration of acetone increases the absorbance, leaving a net decrease in total absorbance as the acetone increases. It's one of those non-intuitive but real effects you run into sometimes - and one which makes attempts to reconstruct "pure material spectra" by algroithmic means tricky, to say the least. Howard | ||
| David Russell (Russell) |
The test mixture I was referring to was a mixture of liquids. Don't recall the specifics, but there was probably no water in the set to make it "easier". | ||
| Tony Davies (Td) |
Just a quick comment that I made to Howard; you can see this affect on the cover of JNIRS. These are spectra of water/ethanol mixtures. I always knew that it was an interesting picture but had never worked out why. On the acetone/water mixtures from looking at individual spectra, I do not believe that the water spectrum will completely overlap acetone. Would anyone out there like to settle the matter? Tony |