Quantifying stabilizing additive hydrolysis and kinetics through principal component analysis of infrared spectra of cross-linked polyethylene pipe

Deep Learning and Infrared Spectroscopy: Representation Learning with a β-Variational Autoencoder

“…12 We have previously observed these correlated spectral changes in carefully designed accelerated aging experiments of PEX-a pipe and found that they result from the (detrimental) hydrolysis of an ester linkage in a stabilizing additive molecule. 12 In Figure 1b, we show the latent traversal for the second-most informative latent dimension (L2). We note that the traversal is dominated by growth in the 1720 cm −1 ketone carbonyl peak absorbance, which is a major thermooxidative degradation product of polyethylene.…”

“…14−16 For example, although PCA projects the data onto linearly independent bases, the principal components (representations) do not necessarily align with the underlying generative factors, which can make interpretation difficult and limit the information that can be learned from such models. 9,12,14,16 In recognition of this, machine learning and deep learning approaches are increasingly being applied to spectroscopy data. 17−22 In this Letter, we have used a deep learning approach to study the spectroscopic and corresponding chemical changes that occur during the aging of cross-linked polyethylene (PEXa) pipe that is increasingly being used for domestic and industrial water transport and heating.…”

“…The physical significance of the results of a PCA of IR spectra can be interpreted by examining the PC score plot in conjunction with the PC loading vectors. 2−9, 12,13 The PCA is guaranteed to yield linearly independent PC representations of the original data (the first condition for disentanglement), but these PCs do not necessarily align with the generative factors and can often be difficult to interpret (the second condition for disentanglement). 9,12,14 Although the PC loading vectors can be interpreted as difference spectra, understanding their significance can be difficult if they are not aligned to generative factors.…”

“…2−9, 12,13 The PCA is guaranteed to yield linearly independent PC representations of the original data (the first condition for disentanglement), but these PCs do not necessarily align with the generative factors and can often be difficult to interpret (the second condition for disentanglement). 9,12,14 Although the PC loading vectors can be interpreted as difference spectra, understanding their significance can be difficult if they are not aligned to generative factors. 9,12,14 In Figures 2 and 3, we show the PC score plot for the first three PCs of our data set and the corresponding loading vectors for each PC.…”

“…9,12,14 Although the PC loading vectors can be interpreted as difference spectra, understanding their significance can be difficult if they are not aligned to generative factors. 9,12,14 In Figures 2 and 3, we show the PC score plot for the first three PCs of our data set and the corresponding loading vectors for each PC. For comparison to the PC score plot, we show in Figure 2 the 3D representation of the three most informative latent dimensions learned by the β-VAE model.…”