As the scale and variety of plastics produced continue to grow, plastics recycling will require innovative solutions. The industrial state-of-the-art sorting technology, near-infrared (NIR) spectroscopy, as currently used, cannot effectively differentiate polyolefins, the single largest class of polymers by volume. Chemical similarity combined with architectural diversity in polyolefins stymies subclass delineation, such as differentiating low-density polyethylene from high-density polyethylene, due to their spectral similarity and chemical overlap. To address this challenge, we use machine learning (ML) to directly predict density, crystallinity, and short-chain branching from NIR spectra, enabling property-based sorting for more effective recycling. After testing a variety of ML models, we find that partial least squares regression provides high prediction accuracy with model simplicity. Since the resulting model leverages the correlated intensities, we develop a method to enhance interpretability by identifying the most important wavenumbers for property prediction, which we then relate to known polyolefin CH3 NIR vibrational absorption bands. This approach provides a linkage between ML model predictions and the underlying polyolefin chemistry and confirms that our models effectively capture spectrum-structure-property relationships in polyolefins, reinforcing the fundamental role of polymer chain structure in determining properties. These findings significantly contribute to the understanding of polyolefin differentiation using NIR spectroscopy, which could inform future advancements in property-based sorting strategies for plastic recycling efficiency.
Li et al. (2026) studied this question.