Prior-Attention-Mechanism-Based Spectral Identifier for Revealing Hazardous Chemical Additives in Commercial PVC

Abstract Additive components largely determine the environmental and health risks of commercial plastics, but rapid, nondestructive identification of multiple additives in polymer matrices remains challenging. To overcome this challenge, we developed a prior-knowledge-guided triple-scale convolutional neural network (SpecNet) that identified 14 additive classes and two polyvinyl chloride (PVC) types from microconfocal Raman spectra. SpecNet integrated spectral chunking, cross-attention with characteristic peak embedding, and a combined BCE and L1 loss for joint classification and semiquantitative ranking of additive abundances. Trained on 2,200 spectra, the model achieved a 95.03% F1-score on an independent environmental test set (1,000 spectra). Platt scaling reduced the Brier score to 0.03, and Monte Carlo dropout uncertainty was <0.05 when SNR > 10. The model-assisted detection threshold for tris(1-chloro-2-propyl) phosphate (TCPP) reached 2.5 wt %, approximately 2-fold lower than the conventional visual detection limit of Raman spectroscopy (∼5 wt %) by leveraging learned multipeak patterns and prior-knowledge attention to extract weak signals below visual detectability. In zero-shot tests on non-PVC polymers, SpecNet correctly identified additives in 82.8% of cases. For additives in microplastic samples after accelerated photoaging (90 days, ∼15 years natural exposure), it maintained 92.48% identification accuracy by dynamically lowering the detection threshold from 0.5 to 0.005. SpecNet offers an interpretable, field-deployable tool for screening hazardous plastic additives in environmental monitoring, recycling, and risk assessment.