Help with principal component analysis of FTIR spectra

Hello everyone,

I got the task to identify the differences between 3 polymers that were ordered from 3 different producers, but are supposed to be the same product. What are qualitative and quantitative differences.

Currently i have data of FTIR measurements of those 3 polymers that are very similar. In total i have 99 measurements grouped into 3 different types which was already known before measuring. To be more clear i measured the same polymer 33 times for each of the 3 different ones.

Now i want to do a PCA analysis to check what and where the differences between those very similar FTIR spectra are.

I am not an expert so i have some very basic questions:

1. What is the observation and what is the variable in my case? I assume variables are the 99 different measured sampels and the observations are the absorption numbers at each wavenumber from the FTIR data. Is that correct?
2. Which other commands and plots should i use to analyse the data according tto my task? I know there is the pca() function which i can use since i have the statistics toolbox. I think there is a scree plot and some others as well, but im not 100% sure how to do that out of the loadings matrix that i get from the pca() command.
3. If i understood correctly at the end for interpretation i then know which of the wavenumbers result in the differences between the 3 products so that i can then focus on them. But what indicator do i have if those are actually relevant differences that have somehthing to do with the composition of the polymers?
4. My idea was also to use the PCA to check whether all of the samples do end up in three groups, which graph do i use for that?

i attached a file with the data of the whole spectrum and one file with only the fingerpprint region of the spectrums. My main interest is the fingerprint region and that file also contains in the first column the wavenumbers in which the measurements were taken and in the last row the type of polymer the measurement belongs to (1, 2 or 3).

I am happy about any further suggestions or tips.

Thanks for your time!

Best regards Dominik

Answer 1

[the cyclist]

Frankly, I don't see why you would use PCA for this analysis. I expect there are some specialized statistical tests for determining differences between spectra, but I don't know them. You could do a keyword search.

I would start by visualizing the data, and then think more carefully about what specific question you want to try to answer statistically. I'm no expert, but these spectra look quite similar. But, of course, maybe even very small deviations are important to you. You could explore by plotting the mean curves from each producer, and plotting those on top of each other, to see where they deviate.

I'm not really sure you need a statistical test at all, in some ways. Maybe the visualization is enough.

load("Fingerprint.mat","Fingerprint")

producer      = Fingerprint(end,2:end);
wavelength = Fingerprint(1:end-1,1);
response = Fingerprint(1:end-1,2:end);

figure
tiledlayout(2,2)

nexttile
plot(response)
title("all producers")

nexttile
plot(response(:,producer==1))
title("producer 1")

nexttile
plot(response(:,producer==2))
title("producer 2")

nexttile
plot(response(:,producer==3))
title("producer 3")