Improving the Optical Sorting of Polyester Bioplastics via Reflectance Spectroscopy and Machine Learning Classification Techniques

Several examples of bio-based, compostable polymers have
reached
commercialization over the past several decades, showing promise for
addressing the plastics sustainability crisis. Although their composability
minimizes the impact in landfills, employing these materials as single
use plastics on a large scale would require exhaustive amounts of
resources for crop production to meet current plastic demands. While
these thermoplastics are, in principle, mechanically recyclable, the
inability to effectively separate them from poly(ethylene terephthalate)
impedes practical recycling programs that would enable their reuse.
This work explores the potential advancement of optical sorting to
enable the classification of polyester bioplastics. Near-infrared
(NIR) and mid-infrared (MIR) spectral data were collected for over
500 samples to be used for machine learning classification. Four classification
schemes were investigated, including random forest (RF), K nearest
neighbors (kNN), and principal component analysis (PCA) coupled with
both schemes (PCA-RF and PCA-kNN). Prediction accuracies >92% were
demonstrated for both IR regions with the ability to further boost
accuracy to >98% by implementing model confidence thresholds. Exploration
of sample attributes and their impact on material classification revealed
that sample color and opacity have the largest impact on classification
in the NIR region, while the MIR region is unimpacted. Additionally,
feature importance analysis and feature reduction were carried out,
showing that a smaller feature set can be implemented in optical sorters
to more efficiently scan samples using only the most informational
wavelengths. Finally, synthetic Gaussian noise was introduced into
the sample spectra to mimic environmental noise to demonstrate that
the classification models have some tolerance to external noise.