| Author | Message | ||
     Luca Vicenzutto |
Dear all. A simple question to you. In your opinion, using your personal experience, how many spectra you need to create a good calibration? Regards | ||
| Bruce H. Campbell (Campclan) |
This is a simple question. The answer is complicated. Some aspects that in my experience decrease the number of calibration samples necessary are (1) extremely good and precise reference values - this is a rare situation, (2) only a few variables - again a rare situation, (3) linear behavior, (4) the range of reference values is much larger than is the size of the standard deviation of the reference method, (5) the distribution of samples is close to uniform over the reange, and (6) adequate intensities of the bands. The latter point also assumes there is not an extreme amount of overlap of all bands. If those conditions are met, one may be able to use as few as 15 to 20 calibration samples. Also there needs to be at least 5 to 10 validation samples. How often are the above conditions met? Very seldom in my experience. If only 20 calibration samples were used I would set up a schedule of almost continuous verification for the installation of the method. Also I would ensure that those around me understood the calibration may need to be modified. Mathematically, one may be able to use fewer samples but this assumes all variables are known. This is a very rare event. If you are looking for a calibration that is very loose, you may be able to use the minimum number of samples. In this case, I would use the predicted values as indicators, not a final number. Finally, most of the time the calibration samples are what are available and thus do not have a wide range of values, have reference values with understated standard deviations, etc. etc. But don't fall into the trap of making up calibration samples as this means you know ALL the variables that affect the spectra. | ||
| Luca Vicenzutto (Vice) |
I'll explain you one of our approach to the problem. We have some materials made by 2-3 components. I can create a broad range of spectra only using percentage variation of the components of the blend. So, using this approach and a variation range between +2% -2% of every component, I can collect about 12-13 spectra and correlate every spectra with known concentration (in fact I weight components on an analytical balance) Do you think that is a bad way to proceed? | ||
| Tony Davies (Td) |
Luca, Many years ago ( perhaps I should abbreviate this to MYA because I seem to include it in all my replies!). MYA, I did something a bit similar with three plastic films which were being used to make a laminate. [See: Analyst, 110, 643-647 (1985); it was MYA!]. If you have three components that add to 100 percent then if you have three levels: +2%, 0, -2% in two components (the third component level is fixed by the other two) you would generate 9 different mixtures but if you made five levels: +2%, +1%, 0, -1%, -2% you would have 25 mixtures. Then you need to do replications (weighing is good but not error free) so you could generate 75 mixtures for your calibration set. If your "product" is made in the same way i.e just mixing components a, b, c then you should get a resonable calibration. Remember to make a separate validation set of mixtures using, say, +1.6%, +0.6%,-0.6%,-1.6%. If your "product" is actually a, b, c + a filler then you have another level of design and you will need an exprimental design to reduce the number of possible mixtures but that is another question! Hope this helps. Best wishes, Tony |