| Author | Message | ||
     Yves Dauphin (Yves) |
Hello everybody, To obtain a calibration is a lot of work. It is especially long when the NIR spectra come from production equipments because it is very unlikely that one succeeds to convince a factory to perturb its fabrication process so that the calibration will data will cover the range of the constituant values and include the many variations of the process. Thus one relies on what she or he will receive from the production. So immediately a lot of problems are apparent: 1) the calibration is made with one NIR instrument and has to be transferred to an other one (say the first one is broken). 2) the calibration is realized for one production equipment and has to be transferred to an other but similar. 3) on the prodution installation used for the calibration some modification is performed (setup parameter, hardware)or the equipment is just ageing and one had to maintain the calibration 4) a different production is made on the same production equipment as the one used for the calibration and the model must be updated. ... In all these circumstances, one wants to avoid to have to rebuild a complete new calibration. From the user point of view, it is the robustness of the method which is questionned. This is why I would like to start a discussion on these huge and difficult topics. I guess that many people will be interested to share suggestions, experiences and bibliographical references. Dr Yves Dauphin | ||
| Kari Aaljoki (Kallekustaa) |
This is an important issue. I've noticed in practice that only changing from stop-flow to flow conditions may have an small impact to spectra and therefore the models developed using stop-flow spectra are not totally perfect. Thus best possible calibrations can be normally achieved using spectra run on real on-line conditions. | ||
| campclan |
To me, calibrations are both the strengths and weaknesses of chemometric NIR. I think some of the reasons there is difficulty is we tend to try to achieve the best precision we can, even if the situation doesn't call for it. By so doing, we focus on a small volume of calibration space and exclude the surroundings. To illustrate, increasing the robustness of a calibration can be done by introducing a measured amount of noise into the spectra after achieving a "good" calibration. This slightly worsens the precision but increases the robustness. What we need is a calibration procedure that uses a large part of the calibration volume but also focuses small enough that acceptable precision is obtained. I think K-nearest neighbor is a possible way to do this. I do not know of any equipment supplier providing this with their software package. Does anyone know of this? Further, is there any two-dimensional approach that would in essence isolate the peaks of interest from any other interaction? If that were possible in a routine fashion, measuring the peak height would be all the data necessary (somewhat simplistic, but it would be nice). Bruce | ||
| Klaas Faber |
In response to a correspondence on calibration transfer and maintenance, specifically: "Further, is there any two-dimensional approach that would in essence isolate the peaks of interest from any other interaction? If that were possible in a routine fashion, measuring the peak height would be all the data necessary (somewhat simplistic, but it would be nice)." Using data such as HPLC-UV (or GC-MS, EEM fluorescence, LC-LC, LC-IR, GC-IR, etc.), it is possible to calibrate in the presence of unknown interferences. Moreover, only a single calibration sample is required. This property of the two-dimensional data is often referred to as the second-order advantage. Kowalski has been the main advocate of this approach over the years. One could say that with two-dimensional data there is no problem with outliers. Just measure two samples (calibration + unknown) with short time intervals, and you obtain a valid model. This is fundamentally different from NIR where you need a large training set, with all its undesirable consequences such as transfer and maintenance. Of course, there must be a suitable two-dimensional technique. For octane rating, for example, there is none as far as I know. Klaas Faber | ||
| hlmark |
Important caveat for Klass' approach: you have to be VERY sure that all your data is correct, since with only two samples, there is no way to do internal error checking. In this regard, the requirement for many samples is an ADvantage of techniques such as NIR, since erroneous data can be detected through comparson with the rest of the data. Howard | ||
| Klaas Faber |
In response to Howard's caveat: "Important caveat for Klass' approach: you have to be VERY sure that all your data is correct, since with only two samples, there is no way to do internal error checking." In the second-order scenario you get predictions, but also estimates for the pure profiles (spectra, chromatograms). Normally, you know how they should look like: non-negative, unimodal, etc. Often you would have a reference spectrum. If the data is not "correct", that would immediately show up as negative parts in the spectra or spurious shoulders in the chromatograms. It is easy to check this conjecture by corrupting data that is otherwise "correct". Second-order calibration solves the background problem and provides all the diagnostics ("internal error checking") that you don't get with NIR, just because the data has intrinsically richer information. I know that the oldest criticism against second-order calibration (<1980) was about the impossibility to do outlier detection. However, this criticism is not valid, simply because the term outlier is obsolete when going to second-order methods. Unfortunately, there is not a second-order method for each analytical problem. Klaas | ||
| hlmark |
Klass - maybe my lack of knowledge of these second-order methods is making me miss some key point, but I don't see, for example, if your calibration sample actually contains 1% concentration of an analyte, but you erroneously think it contains 2%, how you can prevent all your analyses from giving results that are twice as high as their true values - or how you can detect that situation. Howard | ||
| Klaas Faber |
Howard: I was only referring to a certain mathematical property of second-order data. This property enables one to calibrate in the presence of unknown interferents (so-called second-order advantage over e.g. NIR, which is first-order data). That relates to the message that started this discussion - I believe. You're right of course: if you plug in a number that is twice as high, the prediction will be twice as high and this will not show up in any way. However, consider the following: a second-order model essentially behaves as a univariate zero-intercept model with a single calibration sample. There you will have the same "problem": it is not blunder-proof. (What if there is something wrong with the stock solution?) What can you do agains blunders? Just forget about second-order calibration? Moreover, is NIR blunder-proof? Klaas | ||
| hlmark |
Klass - of course not. I wasn't trying to compare NIR with the second-order methods, or claim that they're "better". I was basically agreeing with you and trying to warn Yves against just those types of blunders that you're telling us about, and that even the second-order methods do not "magically" protect us against. Sometimes there's so much hype about new methods and algorithms that you'd think they're "magic answers" to all chemometric problems! - especially to novice users. We went through this several times in the early days of NIR, with the advent of PCR, then PLS, then ANN. The proponents went so far overboard that I guess I'm still burned - so maybe I'm overcautious. But I think it's still better this way than to simply accept any and every new method that comes along and expect it to do your thinking for you, so when I see a new mthod being proposed, I try to warn the recipient that the standard methods of verifying them still apply and are still needed and should be used. Howard |