Machine Learning Strategies for Improved Polyolefin Sorting Based on Near-Infrared Spectroscopy

Polyolefins are cost-effective and exhibit a myriad of desirable characteristics such as chemical, shock, and impact resistance. While practical for consumer and industrial applications, the stability of polyolefins enables long-term environmental persistence. Recycling and reuse of waste materials would mitigate their long-term environmental impact, but incompatible molecular architectures often result in downgraded material properties for the recyclates. To prevent degradation of the polymers during recycling, the various subclasses of low-density (LD-), high-density (HD-), and linear low-density (LLD-) polyethylene (PE) must be separated from one another as well as from polypropylenes (PP) and other polymers. Unfortunately, the chemical similarities between these polyolefins present a large hurdle to facile separation using common high throughput techniques such as near-visible infrared spectroscopy (NIR). In this work, we explore using machine learning (ML) techniques, coupled with NIR measurements, to enable enhanced sorting of polyolefin species beyond what is possible using current NIR databases. NIR spectra of polyolefins were collected for polyolefins spanning a range of branch content, processing conditions, and additives. Common scattering corrections and preprocessing steps such as multiplicative scatter correction, linear detrending, and Savitzky-Golay filtering were evaluated for their effects on classifier outcomes. Data reduction techniques such as functional principal component analysis (fPCA) and uniform manifold approximation and projection (UMAP) were also investigated for data visualization and classification enhancements. This survey of preprocessing steps and ML algorithm combinations identified multiple data pipelines capable of successfully sorting polyolefin materials. Multiple combinations properly distinguished PP from PE, and separated subclasses of these polyolefins. This work discusses the effects of each data analysis step on the final classification results.