A generalizable time-gated Raman spectroscopy-machine learning workflow for accurate identification of challenging plastics in electronic waste.

Electronic waste (E-waste), the fastest-growing solid waste stream, is projected to reach 82 million metric tons globally by 2030. E-waste plastics are difficult to sort due to diverse polymer compositions and the widespread use of pigments and other additives, which introduce complexity and/or strong fluorescence interference for spectroscopic based methods, especially for Raman spectroscopy. Raman spectroscopy coupled with machine learning (ML) has been explored for plastic identification, but its performance and applicability remain limited by the strong fluorescent backgrounds and reliance on small, instrument-specific datasets. In this study, we developed a generalizable analytical framework for accurate identification of the seven most common e-waste plastics (i.e., ABS, PS, PE, PP, PET, PC, PMMA) by integrating time-gated Raman spectroscopy, ML, and community-derived spectral libraries. Briefly, plastic spectra from open-source and our in-house libraries were standardized and combined to train ML models, which were subsequently validated on external data collected for real e-waste plastics obtained from a recycling facility. This ML-based workflow achieved higher identification accuracy compared with conventional correlation-based methods (i.e., 95% vs 66%), including for the dark plastics (i.e., 84% vs 52%), which have been challenging in previous studies. Beyond the improvement in identification performance, this study established a generalizable framework that enables integration and standardization of heterogeneous community-derived spectra for training robust ML models, providing a scalable, transferable and FAIR data principle-aligned solution for real-world e-waste plastic sorting.