Normalization of LCF coefficients when fitting experimental spectra with theoretical spectra
Hi, I am a theoretical chemistry PhD student working on fitting experimental spectra with computed spectra in order to get a better understanding of the composition. In the LCF process with athena, I have found that when I allow "sum coefficients to 1", I get an unphysical negative coefficient of the last standard- after reading through the mailing list I understood why, and so no longer use that. The sum of my coefficients during the fitting now is close to 1- generally within the range of 0.95-1.1 at the absolute extremes, but more often around ~0.98- ~1.05. In order to compare these coefficients, I renormalize them to 1, so they can represent the fractions of each component present. However, to test this I took the new normalized coefficients and summed up the standards with their respective weights to create this normalized "fit", and found that it is worse than the initial one. Is this something I should be concerned about when reporting the qualitative trends in how the composition changes? And if so, is there a better way to do the fits in a more normalized way? Many thanks, Patricia
Hi Patricia, When you say "worse than the initial one," what do you mean? If it's a "worse" match than nonsensical fit results, but makes some physical and chemical sense, I'd personally argue it's a much better fit ;) What you are doing seems reasonable to me, but I would further recommend keeping your fit component selection as simple as possible. I'm a big proponent of two-component fits, maybe three if there's a really good reason. I feel like a lot of people do multicomponent (3-4 fitting spectra, maybe even more) LCF with the argument that the inclusion of additional components improves the fit statistics...But to me that gives a false sense of certainty that the technique doesn't really allow (at least for the systems I work with). For reporting qualitative trends, it seems like you're on the right track. Perhaps unless it's a qualitative trend of a minor component that might not even be there at all and you're basing your entire argument on it definitely being there, which uhh. Seems to happen a lot. Cheers, Mike
On Mar 11, 2021, at 1:42 AM, Patricia Poths
wrote: Hi,
I am a theoretical chemistry PhD student working on fitting experimental spectra with computed spectra in order to get a better understanding of the composition. In the LCF process with athena, I have found that when I allow "sum coefficients to 1", I get an unphysical negative coefficient of the last standard- after reading through the mailing list I understood why, and so no longer use that. The sum of my coefficients during the fitting now is close to 1- generally within the range of 0.95-1.1 at the absolute extremes, but more often around ~0.98- ~1.05.
In order to compare these coefficients, I renormalize them to 1, so they can represent the fractions of each component present. However, to test this I took the new normalized coefficients and summed up the standards with their respective weights to create this normalized "fit", and found that it is worse than the initial one. Is this something I should be concerned about when reporting the qualitative trends in how the composition changes? And if so, is there a better way to do the fits in a more normalized way?
Many thanks, Patricia _______________________________________________ Ifeffit mailing list Ifeffit@millenia.cars.aps.anl.gov http://millenia.cars.aps.anl.gov/mailman/listinfo/ifeffit Unsubscribe: http://millenia.cars.aps.anl.gov/mailman/options/ifeffit
Dear Patricia, Linear combination fits are a nice tool, but one should handle them with great care - the background subtraction process can introduce significant errors. You have to fine tune the parameters by hand if you are using Athena to get reasonable results - sometimes PCA (principle component analysis) helps to find a reasonable number of components - Athena only shows the fitting errors: please keep in mind, that the error of the complete measurement and data evaluation process is much larger. It makes sense to test the process with known mixtures. - the variation of the of the sum of all components between 98 and 105 % (even with choosing -> sum coefficients to 1) is a problem somewhere inside of Athena. I think only Bruce Ravel can help you there. best regards Stefan
Am 08.03.2021 um 19:27 schrieb Patricia Poths
: Hi,
I am a theoretical chemistry PhD student working on fitting experimental spectra with computed spectra in order to get a better understanding of the composition. In the LCF process with athena, I have found that when I allow "sum coefficients to 1", I get an unphysical negative coefficient of the last standard- after reading through the mailing list I understood why, and so no longer use that. The sum of my coefficients during the fitting now is close to 1- generally within the range of 0.95-1.1 at the absolute extremes, but more often around ~0.98- ~1.05.
In order to compare these coefficients, I renormalize them to 1, so they can represent the fractions of each component present. However, to test this I took the new normalized coefficients and summed up the standards with their respective weights to create this normalized "fit", and found that it is worse than the initial one. Is this something I should be concerned about when reporting the qualitative trends in how the composition changes? And if so, is there a better way to do the fits in a more normalized way?
Many thanks, Patricia _______________________________________________ Ifeffit mailing list Ifeffit@millenia.cars.aps.anl.gov http://millenia.cars.aps.anl.gov/mailman/listinfo/ifeffit Unsubscribe: http://millenia.cars.aps.anl.gov/mailman/options/ifeffit
participants (3)
-
Mangold, Stefan (IPS)
-
Mike Massey
-
Patricia Poths