Real-time classification of waste polyethylene using terahertz and raman spectroscopy integrated machine learning

Abstract

Abstract Efficient polymer classification is vital to recycling, yet conventional methods often struggle to distinguish plastics within the same class because of their similar thermophysical properties. This study presents an in-operation machine-learning (ML) approach combining terahertz time-domain spectroscopy (THz-TDS) and Raman spectroscopy for rapid classification of industrial polyethylene (PE)-rich waste feedstocks. Spectral data from six industry-provided PE-rich materials with differences in additives and physical form were collected to build a dual-modality spectral library. Principal component analysis (PCA) and linear discriminant analysis (LDA) were applied for dimensionality reduction, and the resulting features were used to train polymer-classification models. The combined THz–Raman PCA model achieved 94.44% accuracy, while the same approach with individual THz and Raman-only modes achieved 89.26 and 87.41% accuracy, respectively. Adding LDA after PCA overgeneralized the small dataset, reducing combined and THz-only accuracy by up to 3.14% while improving Raman-only accuracy by 1.85%. The results demonstrate the feasibility and complementary value of integrating surface-sensitive Raman information with bulk-sensitive THz features for distinguishing compositionally similar PE-rich waste feedstocks. The complete process takes approximately 33 s: 30 s for THz, 3 s for Raman, and 0.0014 s for model inference, supporting rapid feedstock screening and quality control.