The pervasive occurrence of microplastics (MPs) in aquatic environments presents growing challenges for environmental monitoring. Conventional MP detection methods often require extensive pretreatment and struggle to differentiate mixed polymer compositions in complex matrices. Here, we present a pretreatment-free analytical method that integrates an electrostatically functionalized surface-enhanced Raman scattering (SERS) substrate with an interpretable deep learning framework. A hierarchically porous gold sponge modified with poly(diallyldimethylammonium chloride) facilitates efficient electrostatic enrichment and size-selective capture of negatively charged MPs, while embedded gold nanoparticles generate plasmonic hotspots for enhanced Raman signal amplification. A modular binary convolutional neural network framework (CNN) employing a one-vs-rest architecture enables accurate and interpretable classification of five representative MPs, i.e., polytetrafluoroethylene, polypropylene, polystyrene, polyvinyl chloride, and polyethylene terephthalate, achieving a precision of 0.9896 within the evaluated Raman data set. Gradient-weighted class activation mapping (Grad-CAM) analysis highlights key Raman bands characteristic of each polymer type, e.g., the C-C vibrations in the benzene ring of PET, supporting the chemical interpretability of the CNN model. The system was validated in urban tap water and natural surface waters, which represent both low-interference and heavy-metal-impacted complex matrices. This integrated platform provides a sensitive and adaptable approach for pretreatment-free identification of MPs in complex water matrices, demonstrating its potential for practical environmental analysis.
Sun et al. (Mon,) studied this question.