Challenges in sorting unknown post-consumer plastics using LIBS combined with machine learning

The rampant use of plastics and insufficient disposal practices have put the environment and ecosystems at risk, necessitating efficient sorting methods for recycling. This study demonstrates the potential of a LIBS–ML system for real-time classification of post-consumer plastics using “real, as-received samples.” Using a rigorous LOSO strategy, the evaluation closely simulated realistic recycling conditions. The ANN achieved excellent accuracy for PET, PS, and PVC, while the classification of the HDPE–LDPE–PP group remains a major challenge due to their very similar chemical structures and the large variations present in such real-world materials. Among various analytical strategies compared, the PCA–ANN provided the highest accuracy (99.9%), and LDA–ANN offered nearly comparable performance (98%) with better computational efficiency. All trained models required <2MB storage and achieved inference times <150ms, demonstrating potential for portable deployment. While many studies report ideal performance for all six plastics, our findings show that HDPE, LDPE, and PP remain difficult to separate, emphasizing the need to strengthen their discrimination.