Deep learning-based hyperspectral imaging for the identification and quantification of multi-component polypropylene blends

Characterization of polymer blends and their additives is an essential step towards improving plastic recycling rates. Hyperspectral imaging (HSI) has previously demonstrated effectiveness for quantification of individual additives in polymer blends. This study proposes a deep learning approach to enable the identification and quantification of multiple additives in polymer blends simultaneously. Polypropylene (PP)-based blends containing polyethylene (PE), talc and phosphorous flame retardant (B631C) were prepared for HSI data acquisition. For the quantification task, the regression model achieved high R² values of 0.829-0.937 with low RMSE values of 1.010-2.737 wt%. For the identification task, the multi-label classification model achieved an accuracy of 0.934 ± 0.048. Threshold sensitivity analysis demonstrated a trade-off between B631C detection sensitivity in low-concentration PP/B631C blends and the false-positive rate for neat PP as the decision threshold was varied. The findings of this study provide a foundation for the further development of deep learning frameworks for advanced polymer waste profiling to support appropriate recycling pathway selection towards closed-loop manufacturing.