Enabling mechanical recycling of plastic bottles with shrink sleeves through near-infrared spectroscopy and machine learning algorithms

Shrink sleeves interfere with the mechanical recycling of plastic bottles because of their poor sortability during near-infrared (NIR)-based sorting. This study aims to identify reasons for their poor sortability and to propose solutions to overcome them. Five machine learning (ML) algorithms (decision tree, random forest, support vector machine, partial least squares, and convolutional neural network) are trained on NIR spectra of sleeved and non-sleeved materials and evaluated on different test datasets to investigate the influences of different product de-signs and post-consumer effects on the classification process. The results show that the sortability of sleeved plastic bottles can be significantly improved by (i) avoiding printing or coloring sleeves black and sleeves with large-area printing and high shrinkage; (ii) adding sleeved bottle spectra to the ML training data; (iii) choosing optimal ML algorithms with suitable hyperparameters. With these improvements, the NIR-based sorting and thus mechanical recycling of sleeved bottles could increase greatly.