Machine‐Learning‐Based Prediction of PP / PE / PS Composition Ratios in

ABSTRACT

Plastic recycling requires reliable quality assessment, particularly for the accurate quantification of polymer composition. In this study, a one‐dimensional convolutional neural network (1D‐CNN) was developed to predict the composition ratios of polypropylene (PP), polyethylene (PE), and polystyrene (PS) from attenuated total reflection Fourier‐transform infrared (ATR‐FTIR) spectra. Mixed PP/PE/PS resins covering a wide range of compositions were prepared to enhance model generalizability, and three spectral regions (650–1700, 2700–3100, and 650–3100 cm ⁻¹ ) were evaluated for 1D‐CNN regression. The results showed that the fingerprint model achieved the best performance (MAE ≈ 0.020) for polymer blends, with estimation errors approaching those of quantitative nuclear magnetic resonance (NMR). Application to 26 commercially available recycled PP samples showed that unknown absorption peaks from additives and degradation products reduced the accuracy of the fingerprint‐region model, whereas the full‐spectrum model remained robust across diverse samples. Gradient‐weighted class activation mapping (Grad‐CAM) visualization indicated that the full spectrum model integrated complementary features from both the fingerprint and high‐wavenumber regions, thereby mitigating the influence of unknown peaks. These results suggest that combining ATR‐FTIR with a 1D‐CNN provides a rapid, non‐destructive, and practical approach for predicting the composition of recycled resins, thereby supporting improved quality assurance and the broader utilization of recycled plastics.