Speaker
Description
Machine learning (ML) and artificial intelligence (AI) approaches are revolutionizing the analysis of large data sets. We are contributing to this effort by applying a deep learning approach to analyzing a very large number of infrared (IR) spectra of polymer cross-linked polyethylene (PEX) pipes that are used in household and industrial applications. This work is at the forefront of the application of ML and AI strategies to the analysis of large data sets, an emerging area at the intersection of physical and data science. As PEX pipes age while in use, they sometimes fail unexpectedly. To gain insight into this phenomenon, we use sophisticated instrumentation, high spatial resolution IR microscopy and atomic force microscopy, accelerated aging, and a deep learning analysis to track the evolution of the polyethylene and the stabilizing additives that are incorporated into the pipes to extend their lifetime. Hyperspectral IR images of pipe cross-sections contain a large amount of spatially resolved information about their chemical composition. However, the analysis of these data for complex heterogeneous systems can be challenging because of spectroscopic and spatial complexity. We implement a deep generative modeling approach using a β-variational autoencoder to learn disentangled representations of the generative factors of variance in our large data set of over 30,000 IR spectra collected on cross-sectional slices of unused virgin and used in-service pipes, both with and without cracks. This approach has allowed us to identify three distinct factors of aging and degradation learned by the model, including one associated with crack formation. We also purposely introduce scratches of adjustable depth along the length of the inner surface of the pipes and then subject the pipes to accelerated aging. Our high-resolution IR images have allowed us to identify a new spectroscopic feature near the apex of the scratches that develops during the aging of the pipes. We identified this new feature, associated with a localized enhancement of carboxylate absorbance, through the observation of a large mean square error when processed with our β-VAE model, demonstrating that deep learning approaches can be used to identify new features in data.
| Keyword-1 | deep learning |
|---|---|
| Keyword-2 | infrared spectroscopy |
| Keyword-3 | polymers |