Improving polyolefin sorting via AI-driven property prediction

Polyolefins are the single largest family of polymers produced today. Despite their chemical similarity, the architectural diversity of these polymers lends them to different applications ranging from milk jugs to trash bags to hip replacements. However, this chemical similarity hampers sorting efforts leading to mixed streams with significantly reduced properties and value due to incompatibility or the need for higher reprocessing temperatures that result in degradation. While near-infrared (NIR) spectroscopy is currently used for real-time, non-destructive sorting by recyclers, it struggles with polyolefins. Here we enable improved sorting that can yield higher value products by using artificial intelligence (AI) to predict the density, crystallinity and short chain branching from NIR signals. To generate our dataset and test the robustness of our models, we perform measurements on commercial polyolefins, as well as blends of high-density polyethylene and either polypropylene or low-density polyethylene. Furthermore, we develop a method to provide interpretability of our AI models in an easily accessible way. Ultimately, our approach can be used to create interpretable predictions for improved polyolefin sorting, bring a circular economy to fruition.